You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@hbase.apache.org by "Allan Yang (JIRA)" <ji...@apache.org> on 2016/07/20 03:12:21 UTC
[jira] [Comment Edited] (HBASE-15213) Fix increment performance
regression caused by HBASE-8763 on branch-1.0
[ https://issues.apache.org/jira/browse/HBASE-15213?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15385242#comment-15385242 ]
Allan Yang edited comment on HBASE-15213 at 7/20/16 3:12 AM:
-------------------------------------------------------------
After a little search, I found out that the thought of removing the WriteQueue is already raised by someone else, but soon denied. It seems that if all operations are synced by the WAL, then indeed we don't need a queue to ensure the monotone increasing of MVCC. But there are other operations which don't sync WAL. So we still need a queue to ensure that. Let me quote two comments in HBASE-8763
{quote}
Jeffrey Zhong added a comment - 03/Dec/13 08:00
Today I had some discussion with Enis Soztutar and Ted Yu on this topic and found it might be possible to handle the JIRA issue in a simpler way. Below are the steps:
1) Memstore insert using long.max as the initial write number
2) append no sync
3) sync
4) update WriteEntry's write number to the sequence number returned from Step 2
5) CompleteMemstoreInsert. In this step, make current read point to be >= the sequence number from Step 2. The reasoning behind this is that once we sync till the sequence number, all changes with small sequence numbers are already synced into WAL. Therefore, we should be able to bump up read number to the last sequence number synced.
Currently, we maintain an internal queue which might defer the read point bump up if transactions complete order is different than that of MVCC internal write queue.
By doing above, it's possible to remove the logics maintaining writeQueue so it means we can remove two locking and one queue loop in write code path. Sounds too good to be true . Let me try to write a quick patch and run it against unit tests to see if the idea could fly.
{quote}
{quote}
Jeffrey Zhong added a comment - 03/Dec/13 11:55
I tried a small patch. Since we support SKIP_WAL model, the MVCC.writeQueue is still needed to main the write order because there is no wal sync operation at all. Also there are quite a few test cases doesn't do appendNosync between mvcc.beginMemstoreInsert and mvcc.completeMemstoreInsert so they are needed to be adjusted. So far I didn't find block issues but still need to verify it thoroughly.
{quote}
was (Author: allan163):
After a little search, I found out that the thought of removing the WriteQueue is already raised by someone else, but soon denied. It seems that if all oprations are synced by the WAL, then indedd we don't need a queue to ensure the monotone increasing of MVCC. But there are other operations which don't sync WAL. So we still need a queue to ensure that. Let me quote two commnet in HBASE-8763
{quote}
Jeffrey Zhong added a comment - 03/Dec/13 08:00
Today I had some discussion with Enis Soztutar and Ted Yu on this topic and found it might be possible to handle the JIRA issue in a simpler way. Below are the steps:
1) Memstore insert using long.max as the initial write number
2) append no sync
3) sync
4) update WriteEntry's write number to the sequence number returned from Step 2
5) CompleteMemstoreInsert. In this step, make current read point to be >= the sequence number from Step 2. The reasoning behind this is that once we sync till the sequence number, all changes with small sequence numbers are already synced into WAL. Therefore, we should be able to bump up read number to the last sequence number synced.
Currently, we maintain an internal queue which might defer the read point bump up if transactions complete order is different than that of MVCC internal write queue.
By doing above, it's possible to remove the logics maintaining writeQueue so it means we can remove two locking and one queue loop in write code path. Sounds too good to be true . Let me try to write a quick patch and run it against unit tests to see if the idea could fly.
{quote}
{quote}
Jeffrey Zhong added a comment - 03/Dec/13 11:55
I tried a small patch. Since we support SKIP_WAL model, the MVCC.writeQueue is still needed to main the write order because there is no wal sync operation at all. Also there are quite a few test cases doesn't do appendNosync between mvcc.beginMemstoreInsert and mvcc.completeMemstoreInsert so they are needed to be adjusted. So far I didn't find block issues but still need to verify it thoroughly.
{quote}
> Fix increment performance regression caused by HBASE-8763 on branch-1.0
> -----------------------------------------------------------------------
>
> Key: HBASE-15213
> URL: https://issues.apache.org/jira/browse/HBASE-15213
> Project: HBase
> Issue Type: Sub-task
> Components: Performance
> Reporter: Junegunn Choi
> Assignee: Junegunn Choi
> Fix For: 1.1.4, 1.0.4
>
> Attachments: 15157v3.branch-1.1.patch, HBASE-15213-increment.png, HBASE-15213.branch-1.0.patch, HBASE-15213.v1.branch-1.0.patch
>
>
> This is an attempt to fix the increment performance regression caused by HBASE-8763 on branch-1.0.
> I'm aware that hbase.increment.fast.but.narrow.consistency was added to branch-1.0 (HBASE-15031) to address the issue and a separate work is ongoing on master branch, but anyway, this is my take on the problem.
> I read through HBASE-14460 and HBASE-8763 but it wasn't clear to me what caused the slowdown but I could indeed reproduce the performance regression.
> Test setup:
> - Server: 4-core Xeon 2.4GHz Linux server running mini cluster (100 handlers, JDK 1.7)
> - Client: Another box of the same spec
> - Increments on random 10k records on a single-region table, recreated every time
> Increment throughput (TPS):
> || Num threads || Before HBASE-8763 (d6cc2fb) || branch-1.0 || branch-1.0 (narrow-consistency) ||
> || 1 | 2661 | 2486 | 2359 |
> || 2 | 5048 | 5064 | 4867 |
> || 4 | 7503 | 8071 | 8690 |
> || 8 | 10471 | 10886 | 13980 |
> || 16 | 15515 | 9418 | 18601 |
> || 32 | 17699 | 5421 | 20540 |
> || 64 | 20601 | 4038 | 25591 |
> || 96 | 19177 | 3891 | 26017 |
> We can clearly observe that the throughtput degrades as we increase the number of concurrent requests, which led me to believe that there's severe context switching overhead and I could indirectly confirm that suspicion with cs entry in vmstat output. branch-1.0 shows a much higher number of context switches even with much lower throughput.
> Here are the observations:
> - WriteEntry in the writeQueue can only be removed by the very handler that put it, only when it is at the front of the queue and marked complete.
> - Since a WriteEntry is marked complete after the wait-loop, only one entry can be removed at a time.
> - This stringent condition causes O(N^2) context switches where n is the number of concurrent handlers processing requests.
> So what I tried here is to mark WriteEntry complete before we go into wait-loop. With the change, multiple WriteEntries can be shifted at a time without context switches. I changed writeQueue to LinkedHashSet since fast containment check is needed as WriteEntry can be removed by any handler.
> The numbers look good, it's virtually identical to pre-HBASE-8763 era.
> || Num threads || branch-1.0 with fix ||
> || 1 | 2459 |
> || 2 | 4976 |
> || 4 | 8033 |
> || 8 | 12292 |
> || 16 | 15234 |
> || 32 | 16601 |
> || 64 | 19994 |
> || 96 | 20052 |
> So what do you think about it? Please let me know if I'm missing anything.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)