nfsd: duplicate reply cache improvements for 3.10
Get series with b4
b4 am [email protected]Versions newest first
- v1 Latest6 patches 0 repliesNew
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/6 | nfsd: eliminate one of the DRC cache searchesfs/ | +19 | 0 replies | |
| 2/6 | nfsd: break out comparator into separate functionfs/ | +35 | 0 replies | |
| 3/6 | nfsd: track memory utilization by the DRCfs/ | +17 | 0 replies | |
| 4/6 | nfsd: add new reply_cache_stats file in nfsdfsfs/ fs/ fs/ | +35 | 0 replies | |
| 5/6 | nfsd: keep stats on worst hash balancing seen so farfs/ | +27 | 0 replies | |
| 6/6 | nfsd: scale up the number of DRC hash buckets with cache sizefs/ | +29 | 0 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [PATCH 0/6] nfsd: duplicate reply cache improvements for 3.10
This patchset replaces the last one I sent that starts with:
[PATCH v2 0/5] nfsd: add a reply_cache_stats file to nfsd for tracking DRC performance
I've done quite a bit of testing over the last month or so (with much help from the RH QA folks), and the changes here are largely due to the results of that testing.
The main changes are:
1/ I've dropped the patch to track the max and average time to search the
cache. What I found was that occasionally the search gets preempted to service an IRQ or bottom half handler, and that would blow the stats out of the water. Since they aren't reliably indicative, it's probably best not to track that for now.
In some of the testing I did, I left this in place and disabled IRQs during the search. That's not a suitable change for upstream but it did help prove the efficacy of some of the later changes.
2/ I've added a patch to scale out the number of hash buckets with the
max size of the cache. This seems to greatly improve performance by simply keeping the hash chain lengths low. I saw about 5-6 fold decrease in the average search time with this change. The downside of course is that we have to allocate that many more buckets up front, but I don't think that's entirely unreasonable.
These patches represent the DRC-related changes that I'd like to see merged in 3.10, possibly sooner if we get any reports of performance regressions in the DRC due to the changes that went into 3.9.
Jeff Layton (6):
nfsd: eliminate one of the DRC cache searches
nfsd: break out comparator into separate function
nfsd: track memory utilization by the DRC
nfsd: add new reply_cache_stats file in nfsdfs
nfsd: keep stats on worst hash balancing seen so far
nfsd: scale up the number of DRC hash buckets with cache size
fs/nfsd/cache.h | 1 + fs/nfsd/nfscache.c | 195 +++++++++++++++++++++++++++++++++++++++-------------- fs/nfsd/nfsctl.c | 9 +++ 3 files changed, 155 insertions(+), 50 deletions(-)
Signature · 1 line
1.7.11.7