Opened at 2007-12-18T22:36:57Z
Last modified at 2010-01-15T20:32:55Z
#236 new enhancement
measure the effect of a more memory-local zfec variant on a webapi server
Reported by: | zooko | Owned by: | zooko |
---|---|---|---|
Priority: | minor | Milestone: | undecided |
Component: | code-encoding | Version: | 0.7.0 |
Keywords: | performance memory zfec large | Cc: | |
Launchpad Bug: |
Description ΒΆ
The preliminary results of the load tests ("the webapi servers are the busy points. Each webapi server is on its own machine. Each one is consuming anywhere from 30% to 70% CPU.") suggest that zfec might be a bottleneck in a complete system (grid plus webapi servers). I have some patches to zfec which constrain its use of memory to be more localized in order to fit into smaller caches.
I will measure whethermany zfec processes running in parallel on one server-sized machine (an Athlon64 2.4 GHz, 512 KiB L2 cache, 4 GiB DDR2 RAM) benefit or suffer from this locality optimization.
Change History (6)
comment:1 Changed at 2008-03-08T00:53:51Z by warner
comment:2 Changed at 2008-06-01T21:05:08Z by warner
- Milestone changed from eventually to undecided
comment:3 Changed at 2009-12-13T03:06:09Z by davidsarah
- Keywords performance memory zfec added; scaling removed
comment:4 Changed at 2009-12-13T03:08:01Z by davidsarah
- Keywords scalability removed
comment:5 Changed at 2009-12-13T05:04:01Z by davidsarah
- Priority changed from major to minor
comment:6 Changed at 2010-01-15T20:32:55Z by davidsarah
- Keywords large added
FYI, the new upload/download status timing pages tell me that for a 5MB 3-of-10 download, zfec is running at 89MBps, whereas AES is running at 5.6MBps. So at least for one-at-a-time downloads, AES is far more of a bottleneck than zfec.
For a machine that is in colo, a 2MB download runs zfec at 587MBps and AES at 12MBps. A 16MB download runs zfec at 320MBps and AES at 12MBps.