hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Nicolas Liochon (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-9775) Client write path perf issues
Date Fri, 18 Oct 2013 20:52:48 GMT

    [ https://issues.apache.org/jira/browse/HBASE-9775?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13799503#comment-13799503
] 

Nicolas Liochon commented on HBASE-9775:
----------------------------------------

I've done various tests, on a much smaller cluster than Elliott's.
I observed better write performances on the 0.96 than 0.94, by about 20% when inserting 100m
of rows from an empty cluster.  There are  around 18 regions at this stage IIRC, so the cluster
size should not matter that much when we start from an empty table. I've inserted around 1b
w/o issue on 0.96.

I haven't compared the number of thread. What I see in .96 is that the actual limit is the
limit per region: there is one thread per client and per server. Once one multi operation
on this server is done, another starts. For this reason, there are little operations on multiple
server: they are not synchronized. In theory this gives better performances; Elliott tests
says the opposite, at least at large scales. At least, I've seen that adding a YCSB client
increase the throughput. It would not be the case if the client was maxing the cluster or
client physical capacity. As well, increasing the max per region helped (by about the same
ratio: 50%). So there is for sure room for improvement here.

I will do the comparison with 0.94 beginning of next week for these points (#thread,  impact
of more clients).  I will as well look at the pure CPU performances of the client. From the
tests so far it seems that we can play with the limits parameters to increase / limit the
throughput. This does not explain the ITBLL failure at all.

BTW, I observed better performances when having 2 YCSB instances vs. a single YCSB with 2
threads. I've seen this as well with the .96 before the AsyncProcess implementation. On a
10 nodes cluster the difference was 30%. I've never done this test w/ the .94.

For the ITBLL I would be interested to see the servers & client logs. The SocketTimeoutException
was strange.

[~jmspaggi] It would be great if you could redo the same tests as the ones you've done a while
ago on HBASE-6295: it could help to see if we have a regression of if it's only a matter a
medium / large cluster... 


> Client write path perf issues
> -----------------------------
>
>                 Key: HBASE-9775
>                 URL: https://issues.apache.org/jira/browse/HBASE-9775
>             Project: HBase
>          Issue Type: Bug
>          Components: Client
>    Affects Versions: 0.96.0
>            Reporter: Elliott Clark
>            Priority: Critical
>         Attachments: Charts Search   Cloudera Manager - ITBLL.png, Charts Search   Cloudera
Manager.png, hbase-9775.patch, job_run.log, short_ycsb.png, ycsb_insert_94_vs_96.png
>
>
> Testing on larger clusters has not had the desired throughput increases.



--
This message was sent by Atlassian JIRA
(v6.1#6144)

Mime
View raw message