[
https://issues.apache.org/jira/browse/HBASE-15594?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15226919#comment-15226919
]
Yu Li commented on HBASE-15594:
-------------------------------
Which version of YCSB you're running sir [~stack]? We happen to do benchmarking
these days comparing 1.1.2 (with miscellaneous fixes on increment as well as
reader/writer rowlock backported) and 0.98.12 using ycsb-0.7.0, and it turns
out the pure read(get) performance declined a lot in our 1.1.2. After days of
debugging (actually just located the root cause few hours ago), we found that
there's a *small but fatal* bug in {{HBaseClient10}} that it initializes one
connection per thread. We could find below codes in {{HBaseClient10#init}},
line 135:
{code}
THREAD_COUNT.getAndIncrement();
synchronized(THREAD_COUNT) {
connection = ConnectionFactory.createConnection(config);
}
{code}
After fixing it in below way, the performance recovered:
{code}
THREAD_COUNT.getAndIncrement();
synchronized(THREAD_COUNT) {
if(connection == null) connection =
ConnectionFactory.createConnection(config);
}
{code}
We were using 4 physical nodes as client, each node ran 8 YCSB process, each
process launched 100 threads. We loaded 100GB data into 3 RS cluster, and then
ran each client to do random get for 30 min, and w/o the fix 1.1.2 performance
is ~20% lower than 0.98.12, which could be reproduced steadily. Hope this
information could help.
> [YCSB] Improvements
> -------------------
>
> Key: HBASE-15594
> URL: https://issues.apache.org/jira/browse/HBASE-15594
> Project: HBase
> Issue Type: Umbrella
> Reporter: stack
> Priority: Critical
>
> Running YCSB and getting good results is an arcane art. For example, in my
> testing, a few handlers (100) with as many readers as I had CPUs (48), and
> upping connections on clients to same as #cpus made for 2-3x the throughput.
> The above config changes came of lore; which configurations need tweaking is
> not obvious going by their names, there were no indications from the app on
> where/why we were blocked or on which metrics are important to consider. Nor
> was any of this stuff written down in docs.
> Even still, I am stuck trying to make use of all of the machine. I am unable
> to overrun a server though 8 client nodes trying to beat up a single node
> (workloadc, all random-read, with no data returned -p readallfields=false).
> There is also a strange phenomenon where if I add a few machines, rather than
> 3x the YCSB throughput when 3 nodes in cluster, each machine instead is doing
> about 1/3rd.
> This umbrella issue is to host items that improve our defaults and noting how
> to get good numbers running YCSB. In particular, I want to be able to
> saturate a machine.
> Here are the configs I'm currently working with. I've not done the work to
> figure client-side if they are optimal (weird is how big a difference
> client-side changes can make -- need to fix this). On my 48 cpu machine, I
> can do about 370k random reads a second from data totally cached in
> bucketcache. If I short-circuit the user gets so they don't do any work but
> return immediately, I can do 600k ops a second but the CPUs are at 60-70%
> only. I cannot get them to go above this. Working on it.
> {code}
> <property>
> <name>
> hbase.ipc.server.read.threadpool.size
> </name>
> <value>48</value>
> </property>
> <property>
> <name>
> hbase.regionserver.handler.count
> </name>
> <value>100</value>
> </property>
> <property>
> <name>
> hbase.client.ipc.pool.size
> </name>
> <value>100</value>
> </property>
> <property>
> <name>
> hbase.htable.threads.max
> </name>
> <value>48</value>
> </property>
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
