Bill Gray wrote:

See attachment with two graphs: (1) cache bandwidth, (2) blowup of sustained memory bandwidth region...

Bill, I had some difficulty with this document under ooffice. A recent version seized and an older version didn't seem to render correctly. Could you export it as pdf? Thanks, -john

- X axis has a log scale

- Light blue line is an older system with 32K L1 and 6M L2 caches

- All other measurements on perf34: 32K L1, 256K L2, 30M L3 caches

- Majority of variation in L1 cache region is from the two guest measurements done with no taskset to a VCPU: yellow and maroon lines. Perhaps this reflects the test bouncing between VCPUs in the guest.

- The sustained memory bandwidth for the guest with no pinning is only 80% of native (maroon line), which motivates more convenient and comprehensive numactl for guests.

- Virtualized bandwidth is otherwise nearly in line with native, which confirms the importance of the virtual CPUID communicating actual native cache sizes to cache-size-aware guest applications, since guest apps could benefit from the full size of the native cache. (Guest was started with "-cpu host", but lscpu in guest showed 4M cache despite actual 30M cache.)

-- john.cooper@redhat.com