[
https://issues.apache.org/jira/browse/HBASE-11323?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14027372#comment-14027372
]
Nick Dimiduk commented on HBASE-11323:
--------------------------------------
+1 for this. I think moving offheap by default is a sound approach, barring any
new revelations re: CMS and/or G1 (cough [~apurtell]). We'll enjoy more
stability and have a less complicated story for folks with gobs of ram.
In this new world, the default RS configuration looks something like 4-6G heap
(how many memstores do you need?) + XXX direct memory. Leave the heap for
memstore and protobuf garbage; only 512m - 1g for index blocks in LruBlockCache
that only evicts from compactions. It could even be simplified to remove the 3
different block priorities.
bq. we allow you to have a table sit in LRUBC
I think this becomes pretty easy if we're assuming a CombinedBlockCache is
always present. It's managing the block placement strategy already, so it can
check the flag and do the right thing.
bq. I have BucketCache deploy about 20% less ops than LRUBC when all incache
and maybe 10% less ops when falling out of cache.
Must be smallish heap. What are your memory allocations here? Have you tried
with 24+G heap? I think you'll see very different numbers.
One improvement for BucketCache is to have HFileBlock's buf refer directly to
the memory in BucketCache instead of copying it over. Might help, if we can
ensure a block isn't evicted out from under us.
> BucketCache all the time!
> -------------------------
>
> Key: HBASE-11323
> URL: https://issues.apache.org/jira/browse/HBASE-11323
> Project: HBase
> Issue Type: Sub-task
> Components: io
> Reporter: stack
> Fix For: 0.99.0
>
>
> One way to realize the parent issue is to just enable bucket cache all the
> time; i.e. always have offheap enabled. Would have to do some work to make
> it drop-dead simple on initial setup (I think it doable).
> So, upside would be the offheap upsides (less GC, less likely to go away and
> never come back because of full GC when heap is large, etc.).
> Downside is higher latency. In Nick's BlockCache 101 there is little to no
> difference between onheap and offheap. In a basic compare doing scans and
> gets -- details to follow -- I have BucketCache deploy about 20% less ops
> than LRUBC when all incache and maybe 10% less ops when falling out of cache.
> I can't tell difference in means and 95th and 99th are roughly same (more
> stable with BucketCache). GC profile is much better with BucketCache -- way
> less. BucketCache uses about 7% more user CPU.
> More detail on comparison to follow.
> I think the numbers disagree enough we should probably do the [~lhofhansl]
> suggestion, that we allow you to have a table sit in LRUBC, something the
> current bucket cache layout does not do.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
