[
https://issues.apache.org/jira/browse/HBASE-17072?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15680085#comment-15680085
]
stack commented on HBASE-17072:
-------------------------------
One thought is to do a more radical HBASE-10676 and cache block headers in a
Map per file keyed by offset. Header on a block is 33 bytes. Add Map overhead.
When we read the Map, it is a single threaded context so no need of synchronize
on the Map. Such a Map would save a seek regardless of what thread asks to read
the next block and it'd save a seek on all subsequent Gets of any block from
the file.
Per Gig of HFile with 64kb hfileblocks, caching all block headers is about .5
MB. Map overheaps would bulk it up. 10 files of 10G each in a region of 10
column families with 1000 regions on a regionserver is too much... 55G if I do
the math right. For something more reasonable, 200 regions with four column
familes with four files of 10G each, thats < 2G or so, not so bad.
We write the index, etc., at write time. We could write a flat array of the
vitals from block header at same time... and stick it on the file as meta info.
If present, use it, else do the expensive Map described above.
> CPU usage starts to climb up to 90-100% when using G1GC
> -------------------------------------------------------
>
> Key: HBASE-17072
> URL: https://issues.apache.org/jira/browse/HBASE-17072
> Project: HBase
> Issue Type: Bug
> Components: Performance, regionserver
> Affects Versions: 1.0.0, 1.2.0
> Reporter: Eiichi Sato
> Attachments: disable-block-header-cache.patch, mat-threadlocals.png,
> mat-threads.png, metrics.png, slave1.svg, slave2.svg, slave3.svg, slave4.svg
>
>
> h5. Problem
> CPU usage of a region server in our CDH 5.4.5 cluster, at some point, starts
> to gradually get higher up to nearly 90-100% when using G1GC. We've also run
> into this problem on CDH 5.7.3 and CDH 5.8.2.
> In our production cluster, it normally takes a few weeks for this to happen
> after restarting a RS. We reproduced this on our test cluster and attached
> the results. Please note that, to make it easy to reproduce, we did some
> "anti-tuning" on a table when running tests.
> In metrics.png, soon after we started running some workloads against a test
> cluster (CDH 5.8.2) at about 7 p.m. CPU usage of the two RSs started to rise.
> Flame Graphs (slave1.svg to slave4.svg) are generated from jstack dumps of
> each RS process around 10:30 a.m. the next day.
> After investigating heapdumps from another occurrence on a test cluster
> running CDH 5.7.3, we found that the ThreadLocalMap contain a lot of
> contiguous entries of {{HFileBlock$PrefetchedHeader}} probably due to primary
> clustering. This caused more loops in
> {{ThreadLocalMap#expungeStaleEntries()}}, consuming a certain amount of CPU
> time. What is worse is that the method is called from RPC metrics code,
> which means even a small amount of per-RPC time soon adds up to a huge amount
> of CPU time.
> This is very similar to the issue in HBASE-16616, but we have many
> {{HFileBlock$PrefetchedHeader}} not only {{Counter$IndexHolder}} instances.
> Here are some OQL counts from Eclipse Memory Analyzer (MAT). This shows a
> number of ThreadLocal instances in the ThreadLocalMap of a single handler
> thread.
> {code}
> SELECT *
> FROM OBJECTS (SELECT AS RETAINED SET OBJECTS value
> FROM OBJECTS 0x4ee380430) obj
> WHERE obj.@clazz.@name =
> "org.apache.hadoop.hbase.io.hfile.HFileBlock$PrefetchedHeader"
> #=> 10980 instances
> {code}
> {code}
> SELECT *
> FROM OBJECTS (SELECT AS RETAINED SET OBJECTS value
> FROM OBJECTS 0x4ee380430) obj
> WHERE obj.@clazz.@name = "org.apache.hadoop.hbase.util.Counter$IndexHolder"
> #=> 2052 instances
> {code}
> Although as described in HBASE-16616 this somewhat seems to be an issue in
> G1GC side regarding weakly-reachable objects, we should keep ThreadLocal
> usage minimal and avoid creating an indefinite number (in this case, a number
> of HFiles) of ThreadLocal instances.
> HBASE-16146 removes ThreadLocals from the RPC metrics code. That may solve
> the issue (I just saw the patch, never tested it at all), but the
> {{HFileBlock$PrefetchedHeader}} are still there in the ThreadLocalMap, which
> may cause issues in the future again.
> h5. Our Solution
> We simply removed the whole {{HFileBlock$PrefetchedHeader}} caching and
> fortunately we didn't notice any performance degradation for our production
> workloads.
> Because the PrefetchedHeader caching uses ThreadLocal and because RPCs are
> handled randomly in any of the handlers, small Get or small Scan RPCs do not
> benefit from the caching (See HBASE-10676 and HBASE-11402 for the details).
> Probably, we need to see how well reads are saved by the caching for large
> Scan or Get RPCs and especially for compactions if we really remove the
> caching. It's probably better if we can remove ThreadLocals without breaking
> the current caching behavior.
> FWIW, I'm attaching the patch we applied. It's for CDH 5.4.5.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
