DateTime.ToString(“o”) allocates 32 objects, #7836

Petermarcu · 2016-10-26T22:55:13Z

Fix for #7815

I'm still working on running the corefx tests on this change and possibly adding some but wanted to start the CR process.

I'm seeing an ~90% reduction in allocation count and ~80% reduction in allocated bytes after this change.

@tarekgh @weshaggard @stephentoub

tarekgh · 2016-10-26T22:59:19Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

@@ -957,12 +949,88 @@ internal static String Format(DateTime dateTime, String format, DateTimeFormatIn
            }

            if (format.Length == 1) {
+                switch (format[0]) {
+                    case 'o':


nit: using if-statement maybe better than switch here as we are testing 2 values.

I will do that, it was another thing I debated and switches seemed to be used so much in this file I went with the style of the file :)

tarekgh · 2016-10-26T23:02:16Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+            if (digit > 1)
+            {
+                Append(builder, val / 10, digit - 1);
+            }


this may be expensive as it will be recursive call. I would do it in a loop better

Yeah, I was torn because its limited in its current use to not being too deep but I can convert to a loop.

Not really happy with my options for a loop here. Will still think about it but may stick with recursion if I can't come up with a better idea for how to do it in a loop cheaply.

Every solution I've come up with for the loop requires more allocations and the use of a stack of some kind to append the digits from most significant to least. Given the max recursive depth here is 7 and its usually 2 I think the recursive solution is fine and the code is much more readable.

here is some implementation suggestion but I'll leave it to you to decide if you think the recursion is better here.

internal static void AppendNumber(StringBuilder builder, long val, int digit) { for (int i = 0; i < digit; i++) { builder.Append('0'); } int index = 1; while (val > 0 && index <= digit) { builder[builder.Length - index] = (char)('0' + (val % 10)); val = val / 10; index++; } index = builder.Length - digit; while (val > 0) { builder.Insert(index, (char)('0' + (val % 10))); val = val / 10; } }

The usage of the string builder is expensive. You are first allocating the string builder (even cached), then in the end you are allocating the string each and every time.

We have run into perf issues here and we solved them in the following manner:
https://ayende.com/blog/169798/excerpts-from-the-ravendb-performance-team-report-dates-take-a-lot-of-time

Note that this code does a single allocation for the string, compute the date parts only once.

In our tests, it was 15% of the runtime of the default impl. And it can probably be improved still.

tarekgh · 2016-10-26T23:03:31Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+            return StringBuilderCache.GetStringAndRelease(result);
+        }
+
+        internal static void Append(StringBuilder builder, long val, int digit)


Append [](start = 29, length = 6)

nit: could you call this AppendNumber instead. it will make easier to not confuse it with the StringBuilder Append

tarekgh · 2016-10-26T23:33:55Z

@Petermarcu other than my minor comments, LGTM.

I'm seeing an ~90% reduction in allocation count and ~80% reduction in allocated bytes after this change.

just curious how did you measure it?

tarekgh · 2016-10-27T00:06:25Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs


            return (FormatCustomized(dateTime, format, dtfi, offset));
        }
-
+
+        internal static string RoundTripFormat(ref DateTime dateTime)


RoundTripFormat [](start = 31, length = 15)

one last thing, you named this method as RoundtripFormat which is same name as the defined string constant. would be better to rename the method to something better (e.g. FormatAsISO8601() or anything you think better).

internal const String RoundtripFormat = "yyyy'-'MM'-'dd'T'HH':'mm':'ss.fffffffK";

good feedback

I measured it using the Perf tools in Visual Studio. I can show you if you'd like to see it.

Petermarcu · 2016-10-27T06:41:29Z

@dotnet-bot test Linux ARM Emulator Cross Debug Build please

Petermarcu · 2016-10-27T06:41:51Z

@dotnet-bot test Linux ARM Emulator Cross Release Build please

stephentoub · 2016-10-27T16:14:07Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

-                case 'o':
-                case 'O':
-                    realFormat = RoundtripFormat;
-                    break;


Removing this doesn't break DateTime.Parse?

Actually this is a good catch. the parsing code use this code path too. through ExpandPredefinedFormat which called from DoStrictParse

stephentoub · 2016-10-27T16:17:25Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs


            return (FormatCustomized(dateTime, format, dtfi, offset));
        }
-
+
+        internal static string RoundTripFormat(ref DateTime dateTime)


Why ref DateTime dateTime rather than just DateTime dateTime?

right, DateTime is long size so it can passed as value

stephentoub · 2016-10-27T16:29:57Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+                    break;
+
+                default:
+                    break;


Nit: this default case could be removed

redknightlois · 2016-10-28T02:24:35Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+                builder[builder.Length - index] = (char)('0' + (val % 10));
+                val = val / 10;
+                index++;
+            }
        }


Given that AppendNumber is used multiple times in the same method, I would look if there is a performance improvement by aggresively inlining AppendNumber. JIT may recognize it and build a single function pushing jump offsets to the stack. That would ensure that the code can avoid the call overhead altogether.

EDIT: If not, it's probably worth to open an issue for the JIT team ;)

Petermarcu · 2016-10-28T17:48:31Z

@tarekgh , Can you take one last look? I updated it to handle the DateTimeOffset failure I was seeing and to leverage an existing method I found while investigating. All other feedback was incorporated.

@redknightlois, I didn't see a noticable difference with and without the attribute. If you want to try and measure the overhead and see if there is something actionable for the JIT team, that would be awesome! For now, I'm going to trust the JIT on this one :)

tarekgh · 2016-10-28T17:49:47Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+                builder[builder.Length - index] = (char)('0' + (val % 10));
+                val = val / 10;
+                index++;
+            }


I am seeing you have removed the second loop. I would suggest we just assert val == 0 if we need this method be reliable in case someone else later use it wrong way

tarekgh · 2016-10-28T17:52:17Z

I put another minor comment, other than that LGTM

Petermarcu · 2016-10-28T18:55:09Z

Thanks. I've addressed all feedback. I'll commit once green.

Petermarcu · 2016-10-28T20:49:01Z

@dotnet-bot test Windows_NT x64 Release please

Petermarcu · 2016-10-28T20:49:25Z

@dotnet-bot test Windows_NT x64 Release Priority 1 Build and Test please

weshaggard · 2016-10-30T03:08:18Z

FYI - We have this same code in corert https://github.com/dotnet/corert/blob/master/src/System.Private.CoreLib/src/System/Globalization/FormatProvider.DateTimeFormat.cs.

jkotas · 2016-10-30T08:14:25Z

Opened dotnet/corert#2109 so that we do not forget to port it.

ayende

There might be a better way to handle this, see:
https://ayende.com/blog/169798/excerpts-from-the-ravendb-performance-team-report-dates-take-a-lot-of-time

ayende · 2016-10-30T08:12:58Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+        internal static string RoundTripFormat(ref DateTime dateTime)
+        {
+            StringBuilder result = StringBuilderCache.Acquire();
+            Append(result, dateTime.Year, 4);


Note that calling .Year, .Month, etc are all doing quite a bit of work, and you are doing this each and every time.
This is computationally expensive, and you can probably avoid doing this if you build the entire thing in one go.

ayende · 2016-10-30T08:15:59Z

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

+            if (digit > 1)
+            {
+                Append(builder, val / 10, digit - 1);
+            }


The usage of the string builder is expensive. You are first allocating the string builder (even cached), then in the end you are allocating the string each and every time.

We have run into perf issues here and we solved them in the following manner:
https://ayende.com/blog/169798/excerpts-from-the-ravendb-performance-team-report-dates-take-a-lot-of-time

Note that this code does a single allocation for the string, compute the date parts only once.

In our tests, it was 15% of the runtime of the default impl. And it can probably be improved still.

Petermarcu added 3 commits October 25, 2016 17:58

Adding lower allocating DateTime.ToString('o')

8d69b5d

removing unreachable switch cases

132e5d1

more cleanup

f49c5d1

dnfclas added the cla-already-signed label Oct 26, 2016

tarekgh reviewed Oct 26, 2016

View reviewed changes

tarekgh reviewed Oct 27, 2016

View reviewed changes

stephentoub reviewed Oct 27, 2016

View reviewed changes

src/mscorlib/src/System/Globalization/DateTimeFormat.cs

break;

default:

break;

Copy link

Member

stephentoub Oct 27, 2016

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: this default case could be removed

Petermarcu added 3 commits October 27, 2016 14:14

updates based on review feedback

262c07a

fixing whitespace

dfe10ff

Fix Casing

bf79a18

redknightlois reviewed Oct 28, 2016

View reviewed changes

fixing DateTimeOffset

1321303

tarekgh reviewed Oct 28, 2016

View reviewed changes

adding assert

ec5af22

Petermarcu merged commit 026beb9 into dotnet:master Oct 29, 2016

Petermarcu deleted the DateTime branch October 29, 2016 04:26

ayende reviewed Oct 30, 2016

View reviewed changes

karelz modified the milestone: 2.0.0 Aug 28, 2017

AlexGhiondea added the netfx-port-consider label Oct 4, 2017

Petermarcu mentioned this pull request Jan 31, 2020

Port: DateTime.ToString(“o”) allocates 32 objects dotnet/runtime#19139

Closed

DateTime.ToString(“o”) allocates 32 objects, #7836

DateTime.ToString(“o”) allocates 32 objects, #7836

Uh oh!

Conversation

Petermarcu commented Oct 26, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tarekgh Oct 26, 2016 • 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!

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!

Choose a reason for hiding this comment

Uh oh!

tarekgh commented Oct 26, 2016

Uh oh!

tarekgh Oct 27, 2016 • 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!

Petermarcu commented Oct 27, 2016

Uh oh!

Petermarcu commented Oct 27, 2016

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!

tarekgh Oct 27, 2016 • 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!

redknightlois Oct 28, 2016 • 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!

Petermarcu commented Oct 28, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tarekgh commented Oct 28, 2016

Uh oh!

Petermarcu commented Oct 28, 2016

Uh oh!

Petermarcu commented Oct 28, 2016

Uh oh!

Petermarcu commented Oct 28, 2016

Uh oh!

weshaggard commented Oct 30, 2016

Uh oh!

jkotas commented Oct 30, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ayende left a comment

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!

tarekgh Oct 26, 2016 •

edited

Loading

tarekgh Oct 27, 2016 •

edited

Loading

tarekgh Oct 27, 2016 •

edited

Loading

redknightlois Oct 28, 2016 •

edited

Loading

jkotas commented Oct 30, 2016 •

edited

Loading