On 2017年12月15日 17:06, Martin Kletzander wrote:
On Thu, Dec 14, 2017 at 07:46:27PM +0800, Eli wrote:
@Eli: Can you help with the testing?
It seems the interface is only implement the isolated case, I remember that you have proposed that for some overlap case?
Hi, yes. It got a bit more complicated so I want to do this incrementally. First enable the easiest cases, then add APIs to manage the system's default group, make type='both' allocation work on CDP-enabled hosts, add APIs for modifying cachetunes for live and stopped domains, add support for memory bandwidth allocation, and so on. This is too much stuff to add in one go.
I guess I forgot to add this info to the cover letter (I think I did at least for the previous version).
I also wasted some time on the tests, some of them are not even in the patchset, have a look at previous version if you want to see them.
ok, sorry for not watching for libvirt list for some time..
I have not see the whole patch set yet, but I have some quick testing on you patch, will try to find more time to review patches (Currently I am maintain another daemon software which is dedicated for RDT feature called RMD)
Only the issue 1 is the true issue, for the others, I think they should be discussed, or be treat as the 'known issue'.
My env:
L1d cache: 32K L1i cache: 32K L2 cache: 256K L3 cache: 56320K NUMA node0 CPU(s): 0-21,44-65 NUMA node1 CPU(s): 22-43,66-87
virsh capabilities:
171 <cache> 172 <bank id='0' level='3' type='both' size='55' unit='MiB' cpus='0-21,44-65'> 173 <control granularity='2816' unit='KiB' type='both' maxAllocs='16'/> 174 </bank> 175 <bank id='1' level='3' type='both' size='55' unit='MiB' cpus='22-43,66-87'> 176 <control granularity='2816' unit='KiB' type='both' maxAllocs='16'/> 177 </bank> 178 </cache>
*Issue:
*1. Doesn't support asynchronous cache allocation. e.g, I need provide all cache allocation require ways, but I am only care about the allocation on one of the cache id, cause the VM won't be schedule to another cache (socket).
Oh, really? This is not written in the kernel documentation. Can't the unspecified caches just inherit the setting from the default group? That would make sense. It would also automatically adjust if the default system one is changed.
Maybe I express myself not clearly, yes the caches will be added to default resource group
Do you have any contact to anyone working on the RDT in the kernel? I think this would save time and effort to anyone who will be using the feature. Sure, /fenghua.yu@intel.com and Tony Luck <tony.luck@intel.com>
kernel doc https://github.com/torvalds/linux/blob/master/Documentation/x86/intel_rdt_ui... ///////
So I got this error if I define the domain like this:
<vcpu placement='static'>6</vcpu> <cputune> <emulatorpin cpuset='0,37-38,44,81-82'/> <cachetune vcpus='0-4'> * <cache id='0' level='3' type='both' size='2816' unit='KiB'/> ^^^ not provide cache id='1' * </cachetune>
root@s2600wt:~# virsh start kvm-cat error: Failed to start domain kvm-cat error: Cannot write into schemata file '/sys/fs/resctrl/qemu-qemu-13-kvm-cat-0-4/schemata': Invalid argument
Oh, I have to figure out why is there 'qemu-qemu' :D
This behavior is not correct.
I expect the CBM will be look like:
root@s2600wt:/sys/fs/resctrl# cat qemu-qemu-14-kvm-cat-0-4/* 000000,00000000,00000000 L3:0=80;1=fffff *(no matter what it is, cause my VM won't be schedule on it, ether I have deinfe the vcpu->cpu pining or, I assume that kernel won't schedule it to cache 1)
Well, it matters. It would have to have all zeros there so that that part of the cache is not occupied.
Well, the hardware won't allow you to specify 0 ways , at least 1 (some of the platform it's 2 ways) From my previous experence, I set it to fffff (it will be treat as 0 in the code) it's decided by min_cbm_bits see https://github.com/torvalds/linux/blob/master/Documentation/x86/intel_rdt_ui...
*Or at least, restrict xml when I define this domain, tell me I need to provide all cache ids (even if I have 4 cache but I only run my VM on 'cache 0') *
We could do that. It would allow us to make this better (or lift the restriction) in case this is "fixed" in the kernel.
Or at least in the future we could do this to meet the users half-way:
- Any vcpus that have cachetune enabled for them must also be pinned
- Users need to specify allocations for all cache ids that the vcpu might run on (according to the pinning acquired from before), for all others we'd just simply set it to all zeros or the same bitmask as the system's default group.
But for now we could just copy the system's setting to unspecified caches or request the user to specify everything.
*2. cache way fragment (no good answers)
I see that for now we allocate cache ways start from the low bits, newly created VM will allocate cache from the next way, if some of the VM (allocated ways in the middle, eg it's schemata is 00100) destroyed, and that slot (1 cache way) may not fit others and it will be wasted, But, how can we handle this, seems no good way, rearrangement? That will lead cache missing in a time window I think.
Avoiding fragmentation is not a simple thing. It's impossible to do without any moving, which might be unwanted. This will be solved by providing an API that will tell you move the allocation if you so desire. For now I at least try allocating the smallest region into which the requested allocation fits, so that the unallocated parts are as big as possible.
Agree
3. The admin/user should manually operate the default resource group, that's is to say, after resctrl is mounted, the admin/user should manually change the schemata of default group. Will libvirt provide interface/API to handle it?
Yes, this is planned.
4. Will provide some APIs like `FreeCacheWay` to end user to see how many cache ways could be allocated on the host?
Yes, this should be provided by an API as well.
For other users/orchestrator (nova), they may need to know if a VM can schedule on the host, but the cache ways is not liner, it may have fragment.
5, What if other application want to have some shared cache ways with some of the VM? Libvirt for now try to read all of the resource group (instead of maintain the consumed cache ways itself), so if another resource group was created under /sys/fs/resctl, and the schemata of it is "FFFFF", then libvirt will report not enough room for new VM. But the user actually want to have another Appliation(e.g. ovs, dpdk pmds) share cache ways with the VM created by libvirt.
Adding support for shared allocations is planned as I said before, however this is something that will be needed to be taken care of differently anyway. I don't know how specific the use case would be, but let's say you want to have 8 cache ways allocated for the VM, but share only 4 of them with some DPDK PMD. You can't use "shared" because that would just take some 8 bits even when some of them might be shared with the system's default group. Moreover it means that the allocation can be shared with machines ran in the future. So in this case you need to have the 8 bits exclusively allocated and then (only after the machine is started) pin the PMD process to those 4 cache ways.
For the missing issue from the other email:
If the host enabled CDP, which is to see the host will report l3 cache type code and data. when user don't want code/data cache ways allocated separated, for current implement, it will report not support `both` type l3 cache.
But we can improve this as make code and data schemata the same e.g, if host enabled CDP, but user request 2 `both` type l3 cache.
We can write the schemata looks like:
L3DATA:0=3 L3CODE:0=3
Yes, that's what we want to achieve, but again, in a future patchset.
Hope that answers your questions. Thanks for trying it out, it is really complicated to develop something like this without the actual hardware to test it on. Yep.
Have a nice day, Martin