You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
As discussed w/ Peter and @mmisono , the current 'lack' of overhead gap is due to virtio-blk backend use.
as shown in Spool: major differences in virtio-blk and SPDK performance!
virtio-blk is presumably a botteleneck when it comes to performance differences
hence: performance differences are being watered out!
ISSUE:
we can't directly compare native VM w/ SPDK and CVM w/ SPDK
we can't confirm bottleneck in this way
INSTEAD:
We can either:
i. do measurements to show that virtio-blk is indeed taking up most of the performance
ii. or: compare virito-blk w/ spdk directly -> argue that perf difference shows bottleneck
either i or ii, we need to show that CVM drawbacks make a large difference when SPDK is used
I reran swiotlb benchmarks w/ native (unmodified, diff branch) kernel, to ensure we didn't make a mistake.
Saw minimal perf difference, so current measurements are correct.
Motivate CVM IO
Machines:
The text was updated successfully, but these errors were encountered: