Optimize SqlClient primitive type writes #34048

Wraith2 · 2018-12-12T22:46:15Z

When writing some primitive types to the tds packet stream small arrays are allocated and passed to the WriteByteArray function. I have changed WriteByteArray to WriteBytes and added a span parameter so that either a span or an array can be passed in (but not both, the priority is explained in the comments). The function will attempt to write using only a span unless it encounters an end of packet and needs to use an async continuation, in that case it will allocate an array and copy the remainder of the span into it. For primitive types this will rarely be needed.

These changes can be used to allow the new BitConverter.TryWriteBytes overload with stackalloced buffers to remove memory allocations for writing of 16 and 32 bit reals. This can be extended to guids at a later time.

This PR was split from #32811 and into smaller commits for easier review. It has passed the manual and efcore tests in native mode, the tests cannot be successfully run in managed mode due to #33930 . Performance improvements for this and related PR's are in the original discussion.
/cc @keeratsingh @afsanehr @saurabh500 @David-Engel

…ble path with fallback to array on continuation add WriteByteSpan and WriteByteArray public methods

… which avoid byte array allocation in favour of span

saurabh500 · 2018-12-14T01:42:37Z

src/System.Data.SqlClient/src/System/Data/SqlClient/TdsParserStateObject.cs

+                            {
+                                byte[] tempArray = new byte[len];
+                                Span<byte> copyTempTo = tempArray.AsSpan();
+                                ReadOnlySpan<byte> copyTempFrom = b.Slice(remainder, len);


I am confused. Why do you need b.Slice again here? b.Slice(remainder) was already called on line 3062

Good question. The len needs to be there so that we only copy data not empty buffer in the case where a small part of an array is used (pool rental for example). The remainder should be changed to 0 though. I've changed it and I'm re-running the test suites.

manual and efcore tests all pass with the push from earlier today.

Is this testing status for Windows only or both Windows and Linux?

windows/native, this bit is shared though. I can force it into managed mode and test that way as well.

OK. I will run the Unix validations.

…ready advanced

saurabh500 · 2018-12-14T18:58:51Z

@dotnet-bot Test Windows x86 Release Build Please
Failure in System.Runtime.Tests

01:00 != 00:00:00, dn:(UTC+01:00) Casablanca, sn:(UTC+01:00) Casablanca
Expected: True
Actual: False

   at System.Tests.TimeZoneInfoTests.TimeZoneInfo_DisplayNameStartsWithOffset() in D:\j\workspace\windows-TGrou---f8ac6754\src\System.Runtime\tests\System\TimeZoneInfoTests.cs:line 2227

https://mc.dot.net/#/user/Wraith2/pr~2Fjenkins~2Fdotnet~2Fcorefx~2Fmaster~2F/test~2Ffunctional~2Fcli~2F/46d98751a960f8363255da765ff76a58bdba68d9/workItem/System.Runtime.Tests/analysis/xunit/System.Tests.TimeZoneInfoTests~2FTimeZoneInfo_DisplayNameStartsWithOffset

saurabh500 · 2018-12-14T19:12:37Z

src/System.Data.SqlClient/src/System/Data/SqlClient/TdsParserStateObject.cs

+                            {
+                                byte[] tempArray = new byte[len];
+                                Span<byte> copyTempTo = tempArray.AsSpan();
+                                ReadOnlySpan<byte> copyTempFrom = b.Slice(0, len);


Can't you just get rid of the call at 3048. It doesn't seem to be doing anything useful. And call b.Slice(remainder, len) like you were doing earlier, at this location? if ( array == null) is the only location where you are going to get the rest of the content of b
Does this make sense ?

No, there is no guarantee that the WritePacket call will return a task so it's possible to go around the loop without allocating the array. This was a problem that occurred with one of the SplitPacket tests that was pointed out on the original PR.

OK, now I understand. In that case, what is the effect of just doing a b = b.Slice(remainder, len) at 3048 and eliminate this slicing at 3066. Would that work? Then this line would read ReadOnlySpan<byte> copyTempFrom = b;

saurabh500

Merge pending product team validations.

saurabh500 · 2018-12-14T21:37:54Z

@dotnet-bot Test Packaging All Configurations x64 Debug Build please

saurabh500 · 2018-12-15T02:52:47Z

I am facing some issues with EF tests failure with this change. I don't think this PR caused it but I wanna be sure.
I will keep you posted.

Wraith2 · 2018-12-15T11:03:57Z

Managed or native? and what are you seeing fail?

saurabh500 · 2018-12-17T19:20:45Z

LGTM. Tested Manual Tests on Ubuntu,
EFCore Ubuntu with MARS enabled
EFCore Ubuntu with Rando, MARS.

* change public WriteArrayBytes to private WriteBytes and add span capable path with fallback to array on continuation add WriteByteSpan and WriteByteArray public methods * add netcoreapp specific implementations of WriteFloat and WriteDouble which avoid byte array allocation in favour of span * fixup array compy slice to not use remainder as start since it has already advanced * remove additional span Commit migrated from dotnet/corefx@245d532

Wraith2 added 2 commits December 11, 2018 23:16

change public WriteArrayBytes to private WriteBytes and add span capa…

ae678cf

…ble path with fallback to array on continuation add WriteByteSpan and WriteByteArray public methods

add netcoreapp specific implementations of WriteFloat and WriteDouble…

a543f92

… which avoid byte array allocation in favour of span

Wraith2 changed the title ~~Optimize SqlClient primative type writes~~ Optimize SqlClient primitive type writes Dec 12, 2018

Wraith2 mentioned this pull request Dec 12, 2018

SqlClient memory performance changes #32811

Closed

saurabh500 added the area-System.Data.SqlClient label Dec 13, 2018

saurabh500 reviewed Dec 14, 2018

View reviewed changes

Wraith2 added 2 commits December 14, 2018 10:45

fixup array compy slice to not use remainder as start since it has al…

6062aab

…ready advanced

Merge remote-tracking branch 'dotnet/master' into sqlperf-writespan

46d9875

saurabh500 reviewed Dec 14, 2018

View reviewed changes

remove additional span

0684c90

saurabh500 approved these changes Dec 14, 2018

View reviewed changes

saurabh500 added the * NO MERGE * The PR is not ready for merge yet (see discussion for detailed reasons) label Dec 14, 2018

saurabh500 removed the * NO MERGE * The PR is not ready for merge yet (see discussion for detailed reasons) label Dec 17, 2018

saurabh500 merged commit 245d532 into dotnet:master Dec 17, 2018

Wraith2 deleted the sqlperf-writespan branch December 17, 2018 22:26

karelz added this to the 3.0 milestone Dec 21, 2018

Optimize SqlClient primitive type writes #34048

Optimize SqlClient primitive type writes #34048

Uh oh!

Conversation

Wraith2 commented Dec 12, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

saurabh500 Dec 14, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

saurabh500 commented Dec 14, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

saurabh500 left a comment

Choose a reason for hiding this comment

Uh oh!

saurabh500 commented Dec 14, 2018

Uh oh!

saurabh500 commented Dec 15, 2018

Uh oh!

Wraith2 commented Dec 15, 2018

Uh oh!

saurabh500 commented Dec 17, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

saurabh500 Dec 14, 2018 •

edited

Loading