[
https://issues.apache.org/jira/browse/IMPALA-9134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16977933#comment-16977933
]
ASF subversion and git services commented on IMPALA-9134:
---------------------------------------------------------
Commit 66322f27e36f3c322bfa78726e9742ff110e96fd in impala's branch
refs/heads/master from Tim Armstrong
[ https://gitbox.apache.org/repos/asf?p=impala.git;h=66322f2 ]
IMPALA-9134: pipeline data stream sender flush
This changes the logic so that the last TransmitData
and EndDataStream RPCs can be sent to receivers in
parallel, instead of them being serialized in FlushAndSendEos().
This was a problem because each call to FlushAndSendEos() would
wait for one or two network round-trips, so previously the flush
could easily take sum(2 * RTT). After this change the sending
and waiting for RPCs is pipeline so the time taken should be
closer to 2 * RTT.
Testing:
This code is exercised by existing tests. Ran exhaustive tests.
Perf:
No measurable change on local TPC-H (where network latency isn't a
significant factor).
+----------+-----------------------+---------+------------+------------+----------------+
| Workload | File Format | Avg (s) | Delta(Avg) | GeoMean(s) |
Delta(GeoMean) |
+----------+-----------------------+---------+------------+------------+----------------+
| TPCH(30) | parquet / none / none | 7.64 | -1.37% | 5.00 | +0.29%
|
+----------+-----------------------+---------+------------+------------+----------------+
+----------+----------+-----------------------+--------+-------------+------------+------------+----------------+-------+----------------+---------+-------+
| Workload | Query | File Format | Avg(s) | Base Avg(s) |
Delta(Avg) | StdDev(%) | Base StdDev(%) | Iters | Median Diff(%) | MW Zval |
Tval |
+----------+----------+-----------------------+--------+-------------+------------+------------+----------------+-------+----------------+---------+-------+
| TPCH(30) | TPCH-Q14 | parquet / none / none | 3.75 | 3.39 |
+10.48% | * 19.43% * | 8.88% | 7 | +7.26% | 0.48 |
1.19 |
| TPCH(30) | TPCH-Q13 | parquet / none / none | 13.55 | 12.78 | +6.05%
| 5.43% | 4.75% | 7 | +5.09% | 1.76 | 2.14 |
| TPCH(30) | TPCH-Q2 | parquet / none / none | 1.30 | 1.26 | +3.63%
| 5.53% | 3.27% | 7 | +3.93% | 1.28 | 1.46 |
| TPCH(30) | TPCH-Q12 | parquet / none / none | 2.65 | 2.56 | +3.44%
| 4.23% | 2.92% | 7 | +3.68% | 1.28 | 1.73 |
| TPCH(30) | TPCH-Q15 | parquet / none / none | 3.80 | 3.76 | +0.91%
| 0.81% | 1.80% | 7 | +1.42% | 1.12 | 1.22 |
| TPCH(30) | TPCH-Q8 | parquet / none / none | 5.23 | 5.13 | +2.11%
| 5.41% | 0.48% | 7 | +0.04% | 0.16 | 1.01 |
| TPCH(30) | TPCH-Q10 | parquet / none / none | 5.77 | 5.62 | +2.70%
| 9.84% | 2.79% | 7 | -0.80% | -0.96 | 0.68 |
| TPCH(30) | TPCH-Q21 | parquet / none / none | 28.82 | 28.62 | +0.70%
| 0.59% | 0.37% | 7 | +0.68% | 1.92 | 2.67 |
| TPCH(30) | TPCH-Q6 | parquet / none / none | 1.32 | 1.31 | +0.82%
| 0.36% | 1.53% | 7 | +0.23% | 1.28 | 1.38 |
| TPCH(30) | TPCH-Q11 | parquet / none / none | 1.30 | 1.29 | +0.60%
| 5.11% | 2.94% | 7 | +0.14% | 0.16 | 0.27 |
| TPCH(30) | TPCH-Q19 | parquet / none / none | 4.52 | 4.49 | +0.57%
| 2.98% | 1.58% | 7 | +0.12% | 0.32 | 0.44 |
| TPCH(30) | TPCH-Q16 | parquet / none / none | 2.78 | 2.77 | +0.35%
| 3.21% | 2.02% | 7 | +0.17% | 0.64 | 0.25 |
| TPCH(30) | TPCH-Q3 | parquet / none / none | 5.65 | 5.65 | -0.04%
| 1.62% | 1.91% | 7 | -0.06% | -0.32 | -0.04 |
| TPCH(30) | TPCH-Q1 | parquet / none / none | 10.88 | 10.89 | -0.11%
| 0.76% | 0.90% | 7 | +0.01% | 0.00 | -0.26 |
| TPCH(30) | TPCH-Q4 | parquet / none / none | 2.18 | 2.18 | -0.38%
| 1.33% | 2.48% | 7 | -0.04% | -0.16 | -0.36 |
| TPCH(30) | TPCH-Q5 | parquet / none / none | 3.94 | 3.95 | -0.27%
| 2.37% | 1.33% | 7 | -0.26% | -1.12 | -0.26 |
| TPCH(30) | TPCH-Q7 | parquet / none / none | 22.37 | 22.41 | -0.19%
| 2.94% | 1.92% | 7 | -0.66% | -0.80 | -0.15 |
| TPCH(30) | TPCH-Q22 | parquet / none / none | 3.68 | 3.72 | -1.14%
| 4.95% | 3.99% | 7 | +0.13% | 0.00 | -0.48 |
| TPCH(30) | TPCH-Q20 | parquet / none / none | 2.83 | 2.86 | -1.21%
| 2.69% | 2.20% | 7 | -1.60% | -0.64 | -0.93 |
| TPCH(30) | TPCH-Q17 | parquet / none / none | 4.92 | 5.04 | -2.46%
| 3.11% | 2.63% | 7 | -2.00% | -1.44 | -1.62 |
| TPCH(30) | TPCH-Q9 | parquet / none / none | 16.32 | 17.19 | -5.03%
| 2.76% | 4.56% | 7 | -4.00% | -2.24 | -2.53 |
| TPCH(30) | TPCH-Q18 | parquet / none / none | 20.45 | 23.45 |
-12.82% | 5.49% | * 23.11% * | 7 | -4.89% | -0.96 |
-1.44 |
+----------+----------+-----------------------+--------+-------------+------------+------------+----------------+-------+----------------+---------+-------+
Change-Id: If66add99e4370f8faf22761b336205cc9f9f1867
Reviewed-on: http://gerrit.cloudera.org:8080/14717
Reviewed-by: Tim Armstrong <[email protected]>
Tested-by: Impala Public Jenkins <[email protected]>
> Parallelise flush in data stream sender
> ---------------------------------------
>
> Key: IMPALA-9134
> URL: https://issues.apache.org/jira/browse/IMPALA-9134
> Project: IMPALA
> Issue Type: Sub-task
> Components: Distributed Exec
> Reporter: Tim Armstrong
> Assignee: Tim Armstrong
> Priority: Major
> Labels: perf, scalability
>
> The data stream sender currently does a synchronous RPC to close each channel
> https://github.com/apache/impala/blob/d4648e8/be/src/runtime/krpc-data-stream-sender.cc#L565.
> This is suboptimal because it serializes the network round-trips and takes
> sum(RTT) over all the destinations in the best case, where no data needs to
> be flushed or 2 * sum(RTT) in the worst case if all channels need to flush
> data.
> If the RPCs were done asynchronously and overlapped with each other, we could
> get this down to 2 * max(RTT).
> I'm including this as a subtask of multi-threading because this is going to
> scale poorly as the number of fragment instances increases.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
