On 3/16/22 1:17 PM, Claudio Fontana wrote: > On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>> the first user is the qemu driver, >>>>> >>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>> >>>>> This improves the situation by 400%. >>>>> >>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>> compared with direct qemu migration to a nc socket to a file. >>>>> >>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>> --- >>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>> src/util/virfile.c | 12 ++++++++++++ >>>>> src/util/virfile.h | 1 + >>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>> >>>>> Hello, I initially thought this to be a qemu performance issue, >>>>> so you can find the discussion about this in qemu-devel: >>>>> >>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>> >>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html

Current results show these experimental averages maximum throughput migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP "query-migrate", tests repeated 5 times for each). VM Size is 60G, most of the memory effectively touched before migration, through user application allocating and touching all memory with pseudorandom data. 64K: 5200 Mbps (current situation) 128K: 5800 Mbps 256K: 20900 Mbps 512K: 21600 Mbps 1M: 22800 Mbps 2M: 22800 Mbps 4M: 22400 Mbps 8M: 22500 Mbps 16M: 22800 Mbps 32M: 22900 Mbps 64M: 22900 Mbps 128M: 22800 Mbps This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper.

As for the theoretical limit for the libvirt architecture, I ran a qemu migration directly issuing the appropriate QMP commands, setting the same migration parameters as per libvirt, and then migrating to a socket netcatted to /dev/null via {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : QMP: 37000 Mbps

So although the Pipe size improves things (in particular the large jump is for the 256K size, although 1M seems a very good value), there is still a second bottleneck in there somewhere that accounts for a loss of ~14200 Mbps in throughput.

On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: > On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>> the first user is the qemu driver, >>>>>>> >>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>> >>>>>>> This improves the situation by 400%. >>>>>>> >>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>> >>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>> --- >>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>> src/util/virfile.h | 1 + >>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>> >>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>> >>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>> >>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html > > >> Current results show these experimental averages maximum throughput >> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >> "query-migrate", tests repeated 5 times for each). >> VM Size is 60G, most of the memory effectively touched before migration, >> through user application allocating and touching all memory with >> pseudorandom data. >> >> 64K: 5200 Mbps (current situation) >> 128K: 5800 Mbps >> 256K: 20900 Mbps >> 512K: 21600 Mbps >> 1M: 22800 Mbps >> 2M: 22800 Mbps >> 4M: 22400 Mbps >> 8M: 22500 Mbps >> 16M: 22800 Mbps >> 32M: 22900 Mbps >> 64M: 22900 Mbps >> 128M: 22800 Mbps >> >> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. > > Ok, its bouncing around with noise after 1 MB. So I'd suggest that > libvirt attempt to raise the pipe limit to 1 MB by default, but > not try to go higher. > >> As for the theoretical limit for the libvirt architecture, >> I ran a qemu migration directly issuing the appropriate QMP >> commands, setting the same migration parameters as per libvirt, >> and then migrating to a socket netcatted to /dev/null via >> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >> >> QMP: 37000 Mbps > >> So although the Pipe size improves things (in particular the >> large jump is for the 256K size, although 1M seems a very good value), >> there is still a second bottleneck in there somewhere that >> accounts for a loss of ~14200 Mbps in throughput.

> > In the above tests with libvirt, were you using the > --bypass-cache flag or not ? No, I do not. Tests with ramdisk did not show a notable difference for me, but tests with /dev/null were not possible, since the command line is not accepted: # virsh save centos7 /dev/null Domain 'centos7' saved to /dev/null [OK] # virsh save centos7 /dev/null --bypass-cache error: Failed to save domain 'centos7' to /dev/null error: Failed to create file '/dev/null': Invalid argument > > Hopefully use of O_DIRECT doesn't make a difference for > /dev/null, since the I/O is being immediately thrown > away and so ought to never go into I/O cache. > > In terms of the comparison, we still have libvirt iohelper > giving QEMU a pipe, while your test above gives QEMU a > UNIX socket. > > So I still wonder if the delta is caused by the pipe vs socket > difference, as opposed to netcat vs libvirt iohelper code. I'll look into this aspect, thanks!

* Claudio Fontana (cfontana(a)suse.de) wrote: > On 3/17/22 2:41 PM, Claudio Fontana wrote: >> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>> the first user is the qemu driver, >>>>>>>>> >>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>> >>>>>>>>> This improves the situation by 400%. >>>>>>>>> >>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>> >>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>> --- >>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>> >>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>> >>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>> >>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>> >>> >>>> Current results show these experimental averages maximum throughput >>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>> "query-migrate", tests repeated 5 times for each). >>>> VM Size is 60G, most of the memory effectively touched before migration, >>>> through user application allocating and touching all memory with >>>> pseudorandom data. >>>> >>>> 64K: 5200 Mbps (current situation) >>>> 128K: 5800 Mbps >>>> 256K: 20900 Mbps >>>> 512K: 21600 Mbps >>>> 1M: 22800 Mbps >>>> 2M: 22800 Mbps >>>> 4M: 22400 Mbps >>>> 8M: 22500 Mbps >>>> 16M: 22800 Mbps >>>> 32M: 22900 Mbps >>>> 64M: 22900 Mbps >>>> 128M: 22800 Mbps >>>> >>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>> >>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>> not try to go higher. >>> >>>> As for the theoretical limit for the libvirt architecture, >>>> I ran a qemu migration directly issuing the appropriate QMP >>>> commands, setting the same migration parameters as per libvirt, >>>> and then migrating to a socket netcatted to /dev/null via >>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>> >>>> QMP: 37000 Mbps >>> >>>> So although the Pipe size improves things (in particular the >>>> large jump is for the 256K size, although 1M seems a very good value), >>>> there is still a second bottleneck in there somewhere that >>>> accounts for a loss of ~14200 Mbps in throughput. > > > Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, > and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. > > ~50000 mbps qemu to netcat socket to /dev/null > ~35500 mbps virsh save to /dev/null > > Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). It might be closer to RAM or cache bandwidth limited though; for an extra copy.

Dave > Ciao, > > C > >>> >>> In the above tests with libvirt, were you using the >>> --bypass-cache flag or not ? >> >> No, I do not. Tests with ramdisk did not show a notable difference for me, >> >> but tests with /dev/null were not possible, since the command line is not accepted: >> >> # virsh save centos7 /dev/null >> Domain 'centos7' saved to /dev/null >> [OK] >> >> # virsh save centos7 /dev/null --bypass-cache >> error: Failed to save domain 'centos7' to /dev/null >> error: Failed to create file '/dev/null': Invalid argument >> >> >>> >>> Hopefully use of O_DIRECT doesn't make a difference for >>> /dev/null, since the I/O is being immediately thrown >>> away and so ought to never go into I/O cache. >>> >>> In terms of the comparison, we still have libvirt iohelper >>> giving QEMU a pipe, while your test above gives QEMU a >>> UNIX socket. >>> >>> So I still wonder if the delta is caused by the pipe vs socket >>> difference, as opposed to netcat vs libvirt iohelper code. >> >> I'll look into this aspect, thanks! >

On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: ... > I have lots of questions here, and I tried to involve Jiri and Andrea Righi here, who a long time ago proposed a POSIX_FADV_NOREUSE implementation. > > 1) What is the reason iohelper was introduced? > > 2) Was Jiri's comment about the missing linux implementation of POSIX_FADV_NOREUSE? > > 3) if using O_DIRECT is the only reason for iohelper to exist (...?), would replacing it with posix_fadvise remove the need for iohelper? > > 4) What has stopped Andreas' or another POSIX_FADV_NOREUSE implementation in the kernel? For what I remember (it was a long time ago sorry) I stopped to pursue the POSIX_FADV_NOREUSE idea, because we thought that moving to a memcg-based solution was a better and more flexible approach, assuming memcg would have given some form of specific page cache control. As of today I think we still don't have any specific page cache control feature in memcg, so maybe we could reconsider the FADV_NOREUSE idea (or something similar)? Maybe even introduce a separate FADV_<something> flag if we don't want to bind a specific implementation of this feature to a standard POSIX flag (even if FADV_NOREUSE is still implemented as a no-op in the kernel). The thing that I liked about the fadvise approach is its simplicity from an application perspective, because it's just a syscall and that's it, without having to deal with any other subsystems (cgroups, sysfs, and similar). -Andrea

> > Lots of questions.. > > Thanks for all your insight, > > Claudio > >> >> Dave >> >>> Ciao, >>> >>> C >>> >>>>> >>>>> In the above tests with libvirt, were you using the >>>>> --bypass-cache flag or not ? >>>> >>>> No, I do not. Tests with ramdisk did not show a notable difference for me, >>>> >>>> but tests with /dev/null were not possible, since the command line is not accepted: >>>> >>>> # virsh save centos7 /dev/null >>>> Domain 'centos7' saved to /dev/null >>>> [OK] >>>> >>>> # virsh save centos7 /dev/null --bypass-cache >>>> error: Failed to save domain 'centos7' to /dev/null >>>> error: Failed to create file '/dev/null': Invalid argument >>>> >>>> >>>>> >>>>> Hopefully use of O_DIRECT doesn't make a difference for >>>>> /dev/null, since the I/O is being immediately thrown >>>>> away and so ought to never go into I/O cache. >>>>> >>>>> In terms of the comparison, we still have libvirt iohelper >>>>> giving QEMU a pipe, while your test above gives QEMU a >>>>> UNIX socket. >>>>> >>>>> So I still wonder if the delta is caused by the pipe vs socket >>>>> difference, as opposed to netcat vs libvirt iohelper code. >>>> >>>> I'll look into this aspect, thanks! >>>

On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: > On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >> * Claudio Fontana (cfontana(a)suse.de) wrote: >>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>> >>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>> >>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>> >>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>> >>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>> --- >>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>> >>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>> >>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>> >>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>> >>>>> >>>>>> Current results show these experimental averages maximum throughput >>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>> "query-migrate", tests repeated 5 times for each). >>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>> through user application allocating and touching all memory with >>>>>> pseudorandom data. >>>>>> >>>>>> 64K: 5200 Mbps (current situation) >>>>>> 128K: 5800 Mbps >>>>>> 256K: 20900 Mbps >>>>>> 512K: 21600 Mbps >>>>>> 1M: 22800 Mbps >>>>>> 2M: 22800 Mbps >>>>>> 4M: 22400 Mbps >>>>>> 8M: 22500 Mbps >>>>>> 16M: 22800 Mbps >>>>>> 32M: 22900 Mbps >>>>>> 64M: 22900 Mbps >>>>>> 128M: 22800 Mbps >>>>>> >>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>> >>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>> not try to go higher. >>>>> >>>>>> As for the theoretical limit for the libvirt architecture, >>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>> commands, setting the same migration parameters as per libvirt, >>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>> >>>>>> QMP: 37000 Mbps >>>>> >>>>>> So although the Pipe size improves things (in particular the >>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>> there is still a second bottleneck in there somewhere that >>>>>> accounts for a loss of ~14200 Mbps in throughput. >>> >>> >>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>> >>> ~50000 mbps qemu to netcat socket to /dev/null >>> ~35500 mbps virsh save to /dev/null >>> >>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >> >> It might be closer to RAM or cache bandwidth limited though; for an extra copy. > > I was thinking about sendfile(2) in iohelper, but that probably can't work as the input fd is a socket, I am getting EINVAL. > > One thing that I noticed is: > > ommit afe6e58aedcd5e27ea16184fed90b338569bd042 > Author: Jiri Denemark <jdenemar(a)redhat.com&gt; > Date: Mon Feb 6 14:40:48 2012 +0100 > > util: Generalize virFileDirectFd > > virFileDirectFd was used for accessing files opened with O_DIRECT using > libvirt_iohelper. We will want to use the helper for accessing files > regardless on O_DIRECT and thus virFileDirectFd was generalized and > renamed to virFileWrapperFd. > > > And in particular the comment in src/util/virFile.c: > > /* XXX support posix_fadvise rather than O_DIRECT, if the kernel support > * for that is decent enough. In that case, we will also need to > * explicitly support VIR_FILE_WRAPPER_NON_BLOCKING since > * VIR_FILE_WRAPPER_BYPASS_CACHE alone will no longer require spawning > * iohelper. > */ > > by Jiri Denemark. > > I have lots of questions here, and I tried to involve Jiri and Andrea Righi here, who a long time ago proposed a POSIX_FADV_NOREUSE implementation. > > 1) What is the reason iohelper was introduced? With POSIX you can't get sensible results from poll() on FDs associated with plain files. It will always report the file as readable/writable, and the userspace caller will get blocked any time the I/O operation causes the kernel to read/write from the underlying (potentially very slow) storage. IOW if you give QEMU an FD associated with a plain file and tell it to migrate to that, the guest OS will get stalled.

To avoid this we have to give QEMU an FD that is NOT a plain file, but rather something on which poll() works correctly to avoid blocking. This essentially means a socket or pipe FD. Here enters the iohelper - we give QEMU a pipe whose other end is the iohelper. The iohelper suffers from blocking on read/write but that doesn't matter, because QEMU is isolated from this via the pipe. In theory we could just spawn a thread inside libvirtd todo the same as the iohelper, but using a separate helper process is more robust If not using libvirt, you would use QEMU's 'exec:' migration protocol with 'dd' or 'cat' for the same reasons. Libvirt provides the iohelper so we don't have to deal with portibility questions around 'dd' syntax and can add features like O_DIRECT that cat lacks.

> 2) Was Jiri's comment about the missing linux implementation of POSIX_FADV_NOREUSE? > > 3) if using O_DIRECT is the only reason for iohelper to exist (...?), would replacing it with posix_fadvise remove the need for iohelper? We can't remove the iohelper for the reason above.> > 4) What has stopped Andreas' or another POSIX_FADV_NOREUSE implementation in the kernel? With regards, Daniel

On Fri, Mar 25, 2022 at 11:56:44AM +0100, Claudio Fontana wrote: > Thanks Daniel, > > On 3/25/22 11:33 AM, Daniel P. Berrangé wrote: >> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>> >>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>> >>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>> >>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>> >>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>> --- >>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>> >>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>> >>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>> >>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>> >>>>>>> >>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>> through user application allocating and touching all memory with >>>>>>>> pseudorandom data. >>>>>>>> >>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>> 128K: 5800 Mbps >>>>>>>> 256K: 20900 Mbps >>>>>>>> 512K: 21600 Mbps >>>>>>>> 1M: 22800 Mbps >>>>>>>> 2M: 22800 Mbps >>>>>>>> 4M: 22400 Mbps >>>>>>>> 8M: 22500 Mbps >>>>>>>> 16M: 22800 Mbps >>>>>>>> 32M: 22900 Mbps >>>>>>>> 64M: 22900 Mbps >>>>>>>> 128M: 22800 Mbps >>>>>>>> >>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>> >>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>> not try to go higher. >>>>>>> >>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>> >>>>>>>> QMP: 37000 Mbps >>>>>>> >>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>> >>>>> >>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>> >>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>> ~35500 mbps virsh save to /dev/null >>>>> >>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>> >>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>> >>> I was thinking about sendfile(2) in iohelper, but that probably can't work as the input fd is a socket, I am getting EINVAL. >>> >>> One thing that I noticed is: >>> >>> ommit afe6e58aedcd5e27ea16184fed90b338569bd042 >>> Author: Jiri Denemark <jdenemar(a)redhat.com&gt; >>> Date: Mon Feb 6 14:40:48 2012 +0100 >>> >>> util: Generalize virFileDirectFd >>> >>> virFileDirectFd was used for accessing files opened with O_DIRECT using >>> libvirt_iohelper. We will want to use the helper for accessing files >>> regardless on O_DIRECT and thus virFileDirectFd was generalized and >>> renamed to virFileWrapperFd. >>> >>> >>> And in particular the comment in src/util/virFile.c: >>> >>> /* XXX support posix_fadvise rather than O_DIRECT, if the kernel support >>> * for that is decent enough. In that case, we will also need to >>> * explicitly support VIR_FILE_WRAPPER_NON_BLOCKING since >>> * VIR_FILE_WRAPPER_BYPASS_CACHE alone will no longer require spawning >>> * iohelper. >>> */ >>> >>> by Jiri Denemark. >>> >>> I have lots of questions here, and I tried to involve Jiri and Andrea Righi here, who a long time ago proposed a POSIX_FADV_NOREUSE implementation. >>> >>> 1) What is the reason iohelper was introduced? >> >> With POSIX you can't get sensible results from poll() on FDs associated with >> plain files. It will always report the file as readable/writable, and the >> userspace caller will get blocked any time the I/O operation causes the >> kernel to read/write from the underlying (potentially very slow) storage. >> >> IOW if you give QEMU an FD associated with a plain file and tell it to >> migrate to that, the guest OS will get stalled. > > we send a stop command to qemu just before migrating to a file in virsh save though right? > With virsh restore we also first load the VM, and only then start executing it. > > So for virsh save and virsh restore, this should not be a problem? Still we need the iohelper? The same code is used in libvirt for other commands like 'virsh dump' and snapshots, where the VM remains live though. In general I don't think we should remove the iohelper, because QEMU code is written from the POV that the channels honour O_NOBLOCK.

On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: > * Claudio Fontana (cfontana(a)suse.de) wrote: >> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>> the first user is the qemu driver, >>>>>>>>>> >>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>> >>>>>>>>>> This improves the situation by 400%. >>>>>>>>>> >>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>> >>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>> --- >>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>> >>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>> >>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>> >>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>> >>>> >>>>> Current results show these experimental averages maximum throughput >>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>> "query-migrate", tests repeated 5 times for each). >>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>> through user application allocating and touching all memory with >>>>> pseudorandom data. >>>>> >>>>> 64K: 5200 Mbps (current situation) >>>>> 128K: 5800 Mbps >>>>> 256K: 20900 Mbps >>>>> 512K: 21600 Mbps >>>>> 1M: 22800 Mbps >>>>> 2M: 22800 Mbps >>>>> 4M: 22400 Mbps >>>>> 8M: 22500 Mbps >>>>> 16M: 22800 Mbps >>>>> 32M: 22900 Mbps >>>>> 64M: 22900 Mbps >>>>> 128M: 22800 Mbps >>>>> >>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>> >>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>> not try to go higher. >>>> >>>>> As for the theoretical limit for the libvirt architecture, >>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>> commands, setting the same migration parameters as per libvirt, >>>>> and then migrating to a socket netcatted to /dev/null via >>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>> >>>>> QMP: 37000 Mbps >>>> >>>>> So although the Pipe size improves things (in particular the >>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>> there is still a second bottleneck in there somewhere that >>>>> accounts for a loss of ~14200 Mbps in throughput. >> >> >> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >> >> ~50000 mbps qemu to netcat socket to /dev/null >> ~35500 mbps virsh save to /dev/null >> >> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). > > It might be closer to RAM or cache bandwidth limited though; for an extra copy. I was thinking about sendfile(2) in iohelper, but that probably can't work as the input fd is a socket, I am getting EINVAL.

On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: > On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>> >>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>> >>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>> >>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>> >>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>> --- >>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>> >>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>> >>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>> >>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>> >>>>>> >>>>>>> Current results show these experimental averages maximum throughput >>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>> through user application allocating and touching all memory with >>>>>>> pseudorandom data. >>>>>>> >>>>>>> 64K: 5200 Mbps (current situation) >>>>>>> 128K: 5800 Mbps >>>>>>> 256K: 20900 Mbps >>>>>>> 512K: 21600 Mbps >>>>>>> 1M: 22800 Mbps >>>>>>> 2M: 22800 Mbps >>>>>>> 4M: 22400 Mbps >>>>>>> 8M: 22500 Mbps >>>>>>> 16M: 22800 Mbps >>>>>>> 32M: 22900 Mbps >>>>>>> 64M: 22900 Mbps >>>>>>> 128M: 22800 Mbps >>>>>>> >>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>> >>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>> not try to go higher. >>>>>> >>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>> >>>>>>> QMP: 37000 Mbps >>>>>> >>>>>>> So although the Pipe size improves things (in particular the >>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>> there is still a second bottleneck in there somewhere that >>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>> >>>> >>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>> >>>> ~50000 mbps qemu to netcat socket to /dev/null >>>> ~35500 mbps virsh save to /dev/null >>>> >>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>> >>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >> >> I was thinking about sendfile(2) in iohelper, but that probably >> can't work as the input fd is a socket, I am getting EINVAL. > > Yep, sendfile() requires the input to be a mmapable FD, > and the output to be a socket. > > Try splice() instead which merely requires 1 end to be a > pipe, and the other end can be any FD afaik. > > With regards, > Daniel > I did try splice(), but performance is worse by around 500%. It also fails with EINVAL when trying to use it in combination with O_DIRECT. Tried larger and smaller buffers, flags like SPLICE_F_MORE an SPLICE_F_MOVE in any combination; no change, just awful performance.

Here is the code: #ifdef __linux__ +static ssize_t safesplice(int fdin, int fdout, size_t todo) +{ + unsigned int flags = SPLICE_F_MOVE | SPLICE_F_MORE; + ssize_t ncopied = 0; + + while (todo > 0) { + ssize_t r = splice(fdin, NULL, fdout, NULL, todo, flags); + if (r < 0 && errno == EINTR) + continue; + if (r < 0) + return r; + if (r == 0) + return ncopied; + todo -= r; + ncopied += r; + } + return ncopied; +} + +static ssize_t runIOCopy(const struct runIOParams p) +{ + size_t len = 1024 * 1024; + ssize_t total = 0; + + while (1) { + ssize_t got = safesplice(p.fdin, p.fdout, len); + if (got < 0) + return -1; + if (got == 0) + break; + + total += got; + + /* handle last write truncate in direct case */ + if (got < len && p.isDirect && p.isWrite && !p.isBlockDev) { + if (ftruncate(p.fdout, total) < 0) { + return -4; + } + break; + } + } + return total; +} + +#endif Any ideas welcome, Claudio

On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: > On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>> >>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>> >>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>> >>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>> >>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>> --- >>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>> >>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>> >>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>> >>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>> >>>>>>> >>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>> through user application allocating and touching all memory with >>>>>>>> pseudorandom data. >>>>>>>> >>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>> 128K: 5800 Mbps >>>>>>>> 256K: 20900 Mbps >>>>>>>> 512K: 21600 Mbps >>>>>>>> 1M: 22800 Mbps >>>>>>>> 2M: 22800 Mbps >>>>>>>> 4M: 22400 Mbps >>>>>>>> 8M: 22500 Mbps >>>>>>>> 16M: 22800 Mbps >>>>>>>> 32M: 22900 Mbps >>>>>>>> 64M: 22900 Mbps >>>>>>>> 128M: 22800 Mbps >>>>>>>> >>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>> >>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>> not try to go higher. >>>>>>> >>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>> >>>>>>>> QMP: 37000 Mbps >>>>>>> >>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>> >>>>> >>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>> >>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>> ~35500 mbps virsh save to /dev/null >>>>> >>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>> >>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>> >>> I was thinking about sendfile(2) in iohelper, but that probably >>> can't work as the input fd is a socket, I am getting EINVAL. >> >> Yep, sendfile() requires the input to be a mmapable FD, >> and the output to be a socket. >> >> Try splice() instead which merely requires 1 end to be a >> pipe, and the other end can be any FD afaik. >> > > I did try splice(), but performance is worse by around 500%. Hmm, that's certainly unexpected ! > Any ideas welcome, I learnt there is also a newer copy_file_range call, not sure if that's any better. You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just want to copy everything IIRC. With regards, Daniel

On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: > On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>> >>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>> >>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>> >>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>> >>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>> --- >>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>> >>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>> >>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>> >>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>> >>>>>>> >>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>> through user application allocating and touching all memory with >>>>>>>> pseudorandom data. >>>>>>>> >>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>> 128K: 5800 Mbps >>>>>>>> 256K: 20900 Mbps >>>>>>>> 512K: 21600 Mbps >>>>>>>> 1M: 22800 Mbps >>>>>>>> 2M: 22800 Mbps >>>>>>>> 4M: 22400 Mbps >>>>>>>> 8M: 22500 Mbps >>>>>>>> 16M: 22800 Mbps >>>>>>>> 32M: 22900 Mbps >>>>>>>> 64M: 22900 Mbps >>>>>>>> 128M: 22800 Mbps >>>>>>>> >>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>> >>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>> not try to go higher. >>>>>>> >>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>> >>>>>>>> QMP: 37000 Mbps >>>>>>> >>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>> >>>>> >>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>> >>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>> ~35500 mbps virsh save to /dev/null >>>>> >>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>> >>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>> >>> I was thinking about sendfile(2) in iohelper, but that probably >>> can't work as the input fd is a socket, I am getting EINVAL. >> >> Yep, sendfile() requires the input to be a mmapable FD, >> and the output to be a socket. >> >> Try splice() instead which merely requires 1 end to be a >> pipe, and the other end can be any FD afaik. >> > > I did try splice(), but performance is worse by around 500%. Hmm, that's certainly unexpected ! > Any ideas welcome, I learnt there is also a newer copy_file_range call, not sure if that's any better. You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just want to copy everything IIRC. With regards, Daniel

* Claudio Fontana (cfontana(a)suse.de) wrote: > On 3/28/22 10:31 AM, Daniel P. Berrangé wrote: >> On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: >>> On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >>>> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>>>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>>>> --- >>>>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>>>> >>>>>>>>> >>>>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>>>> through user application allocating and touching all memory with >>>>>>>>>> pseudorandom data. >>>>>>>>>> >>>>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>>>> 128K: 5800 Mbps >>>>>>>>>> 256K: 20900 Mbps >>>>>>>>>> 512K: 21600 Mbps >>>>>>>>>> 1M: 22800 Mbps >>>>>>>>>> 2M: 22800 Mbps >>>>>>>>>> 4M: 22400 Mbps >>>>>>>>>> 8M: 22500 Mbps >>>>>>>>>> 16M: 22800 Mbps >>>>>>>>>> 32M: 22900 Mbps >>>>>>>>>> 64M: 22900 Mbps >>>>>>>>>> 128M: 22800 Mbps >>>>>>>>>> >>>>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>>>> >>>>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>>>> not try to go higher. >>>>>>>>> >>>>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>>>> >>>>>>>>>> QMP: 37000 Mbps >>>>>>>>> >>>>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>>>> >>>>>>> >>>>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>>>> >>>>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>>>> ~35500 mbps virsh save to /dev/null >>>>>>> >>>>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>>>> >>>>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>>>> >>>>> I was thinking about sendfile(2) in iohelper, but that probably >>>>> can't work as the input fd is a socket, I am getting EINVAL. >>>> >>>> Yep, sendfile() requires the input to be a mmapable FD, >>>> and the output to be a socket. >>>> >>>> Try splice() instead which merely requires 1 end to be a >>>> pipe, and the other end can be any FD afaik. >>>> >>> >>> I did try splice(), but performance is worse by around 500%. >> >> Hmm, that's certainly unexpected ! >> >>> Any ideas welcome, >> >> I learnt there is also a newer copy_file_range call, not sure if that's >> any better. >> >> You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just >> want to copy everything IIRC. >> >> With regards, >> Daniel >> > > Crazy idea, would trying to use the parallel migration concept for migrating to/from a file make any sense? > > Not sure if applying the qemu multifd implementation of this would apply, maybe it could be given another implementation for "toFile", trying to use more than one cpu to do the transfer? I can't see a way that would help; well, I could if you could somehow have multiple io helper threads that dealt with it. Dave

On 4/5/22 10:35 AM, Dr. David Alan Gilbert wrote: > * Claudio Fontana (cfontana(a)suse.de) wrote: >> On 3/28/22 10:31 AM, Daniel P. Berrangé wrote: >>> On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: >>>> On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >>>>> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>>>>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>>>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>>>>> --- >>>>>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>>>>> >>>>>>>>>> >>>>>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>>>>> through user application allocating and touching all memory with >>>>>>>>>>> pseudorandom data. >>>>>>>>>>> >>>>>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>>>>> 128K: 5800 Mbps >>>>>>>>>>> 256K: 20900 Mbps >>>>>>>>>>> 512K: 21600 Mbps >>>>>>>>>>> 1M: 22800 Mbps >>>>>>>>>>> 2M: 22800 Mbps >>>>>>>>>>> 4M: 22400 Mbps >>>>>>>>>>> 8M: 22500 Mbps >>>>>>>>>>> 16M: 22800 Mbps >>>>>>>>>>> 32M: 22900 Mbps >>>>>>>>>>> 64M: 22900 Mbps >>>>>>>>>>> 128M: 22800 Mbps >>>>>>>>>>> >>>>>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>>>>> >>>>>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>>>>> not try to go higher. >>>>>>>>>> >>>>>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>>>>> >>>>>>>>>>> QMP: 37000 Mbps >>>>>>>>>> >>>>>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>>>>> >>>>>>>> >>>>>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>>>>> >>>>>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>>>>> ~35500 mbps virsh save to /dev/null >>>>>>>> >>>>>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>>>>> >>>>>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>>>>> >>>>>> I was thinking about sendfile(2) in iohelper, but that probably >>>>>> can't work as the input fd is a socket, I am getting EINVAL. >>>>> >>>>> Yep, sendfile() requires the input to be a mmapable FD, >>>>> and the output to be a socket. >>>>> >>>>> Try splice() instead which merely requires 1 end to be a >>>>> pipe, and the other end can be any FD afaik. >>>>> >>>> >>>> I did try splice(), but performance is worse by around 500%. >>> >>> Hmm, that's certainly unexpected ! >>> >>>> Any ideas welcome, >>> >>> I learnt there is also a newer copy_file_range call, not sure if that's >>> any better. >>> >>> You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just >>> want to copy everything IIRC. >>> >>> With regards, >>> Daniel >>> >> >> Crazy idea, would trying to use the parallel migration concept for migrating to/from a file make any sense? >> >> Not sure if applying the qemu multifd implementation of this would apply, maybe it could be given another implementation for "toFile", trying to use more than one cpu to do the transfer? > > I can't see a way that would help; well, I could if you could > somehow have multiple io helper threads that dealt with it. The first issue I encounter here for both the "virsh save" and "virsh restore" scenarios is that libvirt uses fd: migration, not unix: migration. QEMU supports multifd for unix:, tcp:, vsock: as far as I can see. Current save procedure in QMP in short: {"execute":"migrate-set-capabilities", ...} {"execute":"migrate-set-parameters", ...} {"execute":"getfd","arguments":{"fdname":"migrate"}, ...} fd=26 QEMU_MONITOR_IO_SEND_FD: fd=26 {"execute":"migrate","arguments":{"uri":"fd:migrate"}, ...} Current restore procedure in QMP in short: (start QEMU) {"execute":"migrate-incoming","arguments":{"uri":"fd:21"}, ...} Should I investigate changing libvirt to use unix: for save/restore? Or should I look into changing qemu to somehow accept fd: for multifd, meaning I guess providing multiple fd: uris in the migrate command?

Thank you for your help, Claudio

On 4/7/22 3:53 PM, Dr. David Alan Gilbert wrote: > * Claudio Fontana (cfontana(a)suse.de) wrote: >> On 4/5/22 10:35 AM, Dr. David Alan Gilbert wrote: >>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>> On 3/28/22 10:31 AM, Daniel P. Berrangé wrote: >>>>> On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: >>>>>> On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >>>>>>> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>>>>>>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>>>>>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>>>>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>>>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>>>>>>> --- >>>>>>>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>>>>>>> through user application allocating and touching all memory with >>>>>>>>>>>>> pseudorandom data. >>>>>>>>>>>>> >>>>>>>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>>>>>>> 128K: 5800 Mbps >>>>>>>>>>>>> 256K: 20900 Mbps >>>>>>>>>>>>> 512K: 21600 Mbps >>>>>>>>>>>>> 1M: 22800 Mbps >>>>>>>>>>>>> 2M: 22800 Mbps >>>>>>>>>>>>> 4M: 22400 Mbps >>>>>>>>>>>>> 8M: 22500 Mbps >>>>>>>>>>>>> 16M: 22800 Mbps >>>>>>>>>>>>> 32M: 22900 Mbps >>>>>>>>>>>>> 64M: 22900 Mbps >>>>>>>>>>>>> 128M: 22800 Mbps >>>>>>>>>>>>> >>>>>>>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>>>>>>> >>>>>>>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>>>>>>> not try to go higher. >>>>>>>>>>>> >>>>>>>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>>>>>>> >>>>>>>>>>>>> QMP: 37000 Mbps >>>>>>>>>>>> >>>>>>>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>>>>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>>>>>>> >>>>>>>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>>>>>>> ~35500 mbps virsh save to /dev/null >>>>>>>>>> >>>>>>>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>>>>>>> >>>>>>>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>>>>>>> >>>>>>>> I was thinking about sendfile(2) in iohelper, but that probably >>>>>>>> can't work as the input fd is a socket, I am getting EINVAL. >>>>>>> >>>>>>> Yep, sendfile() requires the input to be a mmapable FD, >>>>>>> and the output to be a socket. >>>>>>> >>>>>>> Try splice() instead which merely requires 1 end to be a >>>>>>> pipe, and the other end can be any FD afaik. >>>>>>> >>>>>> >>>>>> I did try splice(), but performance is worse by around 500%. >>>>> >>>>> Hmm, that's certainly unexpected ! >>>>> >>>>>> Any ideas welcome, >>>>> >>>>> I learnt there is also a newer copy_file_range call, not sure if that's >>>>> any better. >>>>> >>>>> You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just >>>>> want to copy everything IIRC. >>>>> >>>>> With regards, >>>>> Daniel >>>>> >>>> >>>> Crazy idea, would trying to use the parallel migration concept for migrating to/from a file make any sense? >>>> >>>> Not sure if applying the qemu multifd implementation of this would apply, maybe it could be given another implementation for "toFile", trying to use more than one cpu to do the transfer? >>> >>> I can't see a way that would help; well, I could if you could >>> somehow have multiple io helper threads that dealt with it. >> >> The first issue I encounter here for both the "virsh save" and "virsh restore" scenarios is that libvirt uses fd: migration, not unix: migration. >> QEMU supports multifd for unix:, tcp:, vsock: as far as I can see. >> >> Current save procedure in QMP in short: >> >> {"execute":"migrate-set-capabilities", ...} >> {"execute":"migrate-set-parameters", ...} >> {"execute":"getfd","arguments":{"fdname":"migrate"}, ...} fd=26 >> QEMU_MONITOR_IO_SEND_FD: fd=26 >> {"execute":"migrate","arguments":{"uri":"fd:migrate"}, ...} >> >> >> Current restore procedure in QMP in short: >> >> (start QEMU) >> {"execute":"migrate-incoming","arguments":{"uri":"fd:21"}, ...} >> >> >> Should I investigate changing libvirt to use unix: for save/restore? >> Or should I look into changing qemu to somehow accept fd: for multifd, meaning I guess providing multiple fd: uris in the migrate command? > > So I'm not sure this is the right direction; i.e. if multifd is the > right answer to your problem. Of course, just exploring the space.

> However, I think the qemu code probably really really wants to be a > socket. Understood, I'll try to bend libvirt to use unix:/// and see how far I get, Thanks, Claudio > > Dave > >> >> Thank you for your help, >> >> Claudio >>

* Claudio Fontana (cfontana(a)suse.de) wrote: > On 4/7/22 3:57 PM, Claudio Fontana wrote: >> On 4/7/22 3:53 PM, Dr. David Alan Gilbert wrote: >>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>> On 4/5/22 10:35 AM, Dr. David Alan Gilbert wrote: >>>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>>> On 3/28/22 10:31 AM, Daniel P. Berrangé wrote: >>>>>>> On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote: >>>>>>>> On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >>>>>>>>> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>>>>>>>>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>>>>>>>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>>>>>>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>>>>>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>>>>>>>>> --- >>>>>>>>>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>>>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>>>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>>>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>>>>>>>>> through user application allocating and touching all memory with >>>>>>>>>>>>>>> pseudorandom data. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>>>>>>>>> 128K: 5800 Mbps >>>>>>>>>>>>>>> 256K: 20900 Mbps >>>>>>>>>>>>>>> 512K: 21600 Mbps >>>>>>>>>>>>>>> 1M: 22800 Mbps >>>>>>>>>>>>>>> 2M: 22800 Mbps >>>>>>>>>>>>>>> 4M: 22400 Mbps >>>>>>>>>>>>>>> 8M: 22500 Mbps >>>>>>>>>>>>>>> 16M: 22800 Mbps >>>>>>>>>>>>>>> 32M: 22900 Mbps >>>>>>>>>>>>>>> 64M: 22900 Mbps >>>>>>>>>>>>>>> 128M: 22800 Mbps >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>>>>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>>>>>>>>> not try to go higher. >>>>>>>>>>>>>> >>>>>>>>>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>>>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>>>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>>>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>>>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> QMP: 37000 Mbps >>>>>>>>>>>>>> >>>>>>>>>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>>>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>>>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>>>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>>>>>>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>>>>>>>>> >>>>>>>>>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>>>>>>>>> ~35500 mbps virsh save to /dev/null >>>>>>>>>>>> >>>>>>>>>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>>>>>>>>> >>>>>>>>>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>>>>>>>>> >>>>>>>>>> I was thinking about sendfile(2) in iohelper, but that probably >>>>>>>>>> can't work as the input fd is a socket, I am getting EINVAL. >>>>>>>>> >>>>>>>>> Yep, sendfile() requires the input to be a mmapable FD, >>>>>>>>> and the output to be a socket. >>>>>>>>> >>>>>>>>> Try splice() instead which merely requires 1 end to be a >>>>>>>>> pipe, and the other end can be any FD afaik. >>>>>>>>> >>>>>>>> >>>>>>>> I did try splice(), but performance is worse by around 500%. >>>>>>> >>>>>>> Hmm, that's certainly unexpected ! >>>>>>> >>>>>>>> Any ideas welcome, >>>>>>> >>>>>>> I learnt there is also a newer copy_file_range call, not sure if that's >>>>>>> any better. >>>>>>> >>>>>>> You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just >>>>>>> want to copy everything IIRC. >>>>>>> >>>>>>> With regards, >>>>>>> Daniel >>>>>>> >>>>>> >>>>>> Crazy idea, would trying to use the parallel migration concept for migrating to/from a file make any sense? >>>>>> >>>>>> Not sure if applying the qemu multifd implementation of this would apply, maybe it could be given another implementation for "toFile", trying to use more than one cpu to do the transfer? >>>>> >>>>> I can't see a way that would help; well, I could if you could >>>>> somehow have multiple io helper threads that dealt with it. >>>> >>>> The first issue I encounter here for both the "virsh save" and "virsh restore" scenarios is that libvirt uses fd: migration, not unix: migration. >>>> QEMU supports multifd for unix:, tcp:, vsock: as far as I can see. >>>> >>>> Current save procedure in QMP in short: >>>> >>>> {"execute":"migrate-set-capabilities", ...} >>>> {"execute":"migrate-set-parameters", ...} >>>> {"execute":"getfd","arguments":{"fdname":"migrate"}, ...} fd=26 >>>> QEMU_MONITOR_IO_SEND_FD: fd=26 >>>> {"execute":"migrate","arguments":{"uri":"fd:migrate"}, ...} >>>> >>>> >>>> Current restore procedure in QMP in short: >>>> >>>> (start QEMU) >>>> {"execute":"migrate-incoming","arguments":{"uri":"fd:21"}, ...} >>>> >>>> >>>> Should I investigate changing libvirt to use unix: for save/restore? >>>> Or should I look into changing qemu to somehow accept fd: for multifd, meaning I guess providing multiple fd: uris in the migrate command? >>> >>> So I'm not sure this is the right direction; i.e. if multifd is the >>> right answer to your problem. >> >> Of course, just exploring the space. > > > I have some progress on multifd if we can call it so: > > I wrote a simple program that sets up a unix socket, > listens for N_CHANNELS + 1 connections there, sets up multifd parameters, and runs the migration, > spawning threads for each incoming connection from QEMU, creating a file to use to store the migration data coming from qemu (optionally using O_DIRECT). > > This program plays the role of a "iohelper"-like thing, basically just copying things over, making O_DIRECT possible. > > I save the data streams to multiple files; this works, for the actual results though I will have to migrate to a better hardware setup (enterprise nvme + fast cpu, under various memory configurations). > > The intuition would be that if we have enough cpus to spare (no libvirt in the picture as mentioned for now), > say, the same 4 cpus already allocated for a certain VM to run, we can use those cpus (now "free" since we suspended the guest) > to compress each multifd channel (multifd-zstd? multifd-zlib?), thus reducing the amount of stuff that needs to go to disk, making use of those cpus. Yes possibly; you have an advantage over ormal migration, in that your vCPUs are stopped.

> Work in progress... > >> >>> However, I think the qemu code probably really really wants to be a >>> socket. >> >> Understood, I'll try to bend libvirt to use unix:/// and see how far I get, >> >> Thanks, >> >> Claudio >> >>> >>> Dave >>> >>>> >>>> Thank you for your help, >>>> >>>> Claudio >>>> >> >

On 3/26/22 4:49 PM, Claudio Fontana wrote: > On 3/25/22 12:29 PM, Daniel P. Berrangé wrote: >> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote: >>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote: >>>> * Claudio Fontana (cfontana(a)suse.de) wrote: >>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote: >>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote: >>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote: >>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote: >>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote: >>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote: >>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote: >>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote: >>>>>>>>>>>>> the first user is the qemu driver, >>>>>>>>>>>>> >>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe size (64k). >>>>>>>>>>>>> >>>>>>>>>>>>> This improves the situation by 400%. >>>>>>>>>>>>> >>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty (~15%-ish) >>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file. >>>>>>>>>>>>> >>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana(a)suse.de&gt; >>>>>>>>>>>>> --- >>>>>>>>>>>>> src/qemu/qemu_driver.c | 6 +++--- >>>>>>>>>>>>> src/qemu/qemu_saveimage.c | 11 ++++++----- >>>>>>>>>>>>> src/util/virfile.c | 12 ++++++++++++ >>>>>>>>>>>>> src/util/virfile.h | 1 + >>>>>>>>>>>>> 4 files changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>>> >>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue, >>>>>>>>>>>>> so you can find the discussion about this in qemu-devel: >>>>>>>>>>>>> >>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)" >>>>>>>>>>>>> >>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html >>>>>>> >>>>>>> >>>>>>>> Current results show these experimental averages maximum throughput >>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP >>>>>>>> "query-migrate", tests repeated 5 times for each). >>>>>>>> VM Size is 60G, most of the memory effectively touched before migration, >>>>>>>> through user application allocating and touching all memory with >>>>>>>> pseudorandom data. >>>>>>>> >>>>>>>> 64K: 5200 Mbps (current situation) >>>>>>>> 128K: 5800 Mbps >>>>>>>> 256K: 20900 Mbps >>>>>>>> 512K: 21600 Mbps >>>>>>>> 1M: 22800 Mbps >>>>>>>> 2M: 22800 Mbps >>>>>>>> 4M: 22400 Mbps >>>>>>>> 8M: 22500 Mbps >>>>>>>> 16M: 22800 Mbps >>>>>>>> 32M: 22900 Mbps >>>>>>>> 64M: 22900 Mbps >>>>>>>> 128M: 22800 Mbps >>>>>>>> >>>>>>>> This above is the throughput out of patched libvirt with multiple Pipe Sizes for the FDWrapper. >>>>>>> >>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that >>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but >>>>>>> not try to go higher. >>>>>>> >>>>>>>> As for the theoretical limit for the libvirt architecture, >>>>>>>> I ran a qemu migration directly issuing the appropriate QMP >>>>>>>> commands, setting the same migration parameters as per libvirt, >>>>>>>> and then migrating to a socket netcatted to /dev/null via >>>>>>>> {"execute": "migrate", "arguments": { "uri", "unix:///tmp/netcat.sock" } } : >>>>>>>> >>>>>>>> QMP: 37000 Mbps >>>>>>> >>>>>>>> So although the Pipe size improves things (in particular the >>>>>>>> large jump is for the 256K size, although 1M seems a very good value), >>>>>>>> there is still a second bottleneck in there somewhere that >>>>>>>> accounts for a loss of ~14200 Mbps in throughput. >>>>> >>>>> >>>>> Interesting addition: I tested quickly on a system with faster cpus and larger VM sizes, up to 200GB, >>>>> and the difference in throughput libvirt vs qemu is basically the same ~14500 Mbps. >>>>> >>>>> ~50000 mbps qemu to netcat socket to /dev/null >>>>> ~35500 mbps virsh save to /dev/null >>>>> >>>>> Seems it is not proportional to cpu speed by the looks of it (not a totally fair comparison because the VM sizes are different). >>>> >>>> It might be closer to RAM or cache bandwidth limited though; for an extra copy. >>> >>> I was thinking about sendfile(2) in iohelper, but that probably >>> can't work as the input fd is a socket, I am getting EINVAL. >> >> Yep, sendfile() requires the input to be a mmapable FD, >> and the output to be a socket. >> >> Try splice() instead which merely requires 1 end to be a >> pipe, and the other end can be any FD afaik. >> >> With regards, >> Daniel >> > > I did try splice(), but performance is worse by around 500%. > > It also fails with EINVAL when trying to use it in combination with O_DIRECT. > > Tried larger and smaller buffers, flags like SPLICE_F_MORE an SPLICE_F_MOVE in any combination; no change, just awful performance. Ok I found a case where splice actually helps: in the read case, without O_DIRECT, splice seems to actually outperform read/write by _a lot_.

I will code up the patch and start making more experiments with larger VM sizes etc. Thanks! Claudio > > Here is the code: > > #ifdef __linux__ > +static ssize_t safesplice(int fdin, int fdout, size_t todo) > +{ > + unsigned int flags = SPLICE_F_MOVE | SPLICE_F_MORE; > + ssize_t ncopied = 0; > + > + while (todo > 0) { > + ssize_t r = splice(fdin, NULL, fdout, NULL, todo, flags); > + if (r < 0 && errno == EINTR) > + continue; > + if (r < 0) > + return r; > + if (r == 0) > + return ncopied; > + todo -= r; > + ncopied += r; > + } > + return ncopied; > +} > + > +static ssize_t runIOCopy(const struct runIOParams p) > +{ > + size_t len = 1024 * 1024; > + ssize_t total = 0; > + > + while (1) { > + ssize_t got = safesplice(p.fdin, p.fdout, len); > + if (got < 0) > + return -1; > + if (got == 0) > + break; > + > + total += got; > + > + /* handle last write truncate in direct case */ > + if (got < len && p.isDirect && p.isWrite && !p.isBlockDev) { > + if (ftruncate(p.fdout, total) < 0) { > + return -4; > + } > + break; > + } > + } > + return total; > +} > + > +#endif > > > Any ideas welcome, > > Claudio >