Batch alloc from bin with behaviour patterned after jemalloc's #334

dsm9000 · 2023-10-24T03:57:16Z

No description provided.

deadalnix · 2023-10-24T22:26:10Z

sdlib/d/gc/arena.d

+		return buffer[0];
+	}
+
+	uint allocSmallBatch(shared(ExtentMap)* emap, size_t size,


Is the batch small? The good news is that it doesn't matter much: this is not claled, you can just tweak allocSmall and be done with it.

deadalnix · 2023-10-24T22:34:34Z

sdlib/d/gc/bin.d

+			auto nFree = slab.freeSlots;
+			if (nFree == 0) {
+				break;
+			}


Why would that happen?

Thinking about it, AFAIK it can't, even on OOM. That part ought to go

deadalnix · 2023-10-24T22:37:34Z

sdlib/d/gc/bin.d

+			auto toFill = needFill - filled;
+			auto count = toFill < nFree ? toFill : nFree;


This is unnecessary.

deadalnix · 2023-10-24T22:39:26Z

sdlib/d/gc/bin.d

+			auto count = toFill < nFree ? toFill : nFree;
+			auto gotSlots =
+				slab.batchAllocate(buffer[filled .. filled + count], size);
+			filled += gotSlots;


filled += slab.batchAllocate(buffer[filled .. buffer.length], size);

deadalnix · 2023-10-24T22:56:21Z

sdlib/d/gc/bin.d

@@ -30,20 +30,31 @@ struct Bin {
 		import d.gc.slab;
 		auto size = binInfos[sizeClass].itemSize;

+		auto needFill = buffer.length;


If you plan on filling the buffer, then the best is likely to slice the remaining part of the buffer at each iterration and check that the size of the buffer is greater than 0 int he loop.

But I'm wondering if we are not better off having a minimal number of allocation requested as a parameter, and possibly a bigger buffer, so that we can fill in on a slab granularity.

The latter strategy is done by other allocators, but not jemalloc, I'm not sure why.

Wouldn't filling in on a slab granularity require a rather large minimal cache bin size? At least for the smaller size classes.

512 items for small size classes do not seem absurd.

So: cache 1 slab's worth per class ? sounds good imho (or maybe 1 page's worth?)

deadalnix · 2023-10-24T22:59:09Z

It be more useful for anyone going over this code that the title describes what this does rather than the fact it does something similar to jemalloc, because then, one has not been enlightened much.

deadalnix · 2023-10-24T23:07:56Z

sdlib/d/gc/bin.d

@@ -153,3 +164,64 @@ private:
 		return current;
 	}
 }
+
+unittest batchAllocation {


This is not testing interesting behavior, such as what happens when there are several partially filled slab in the bin.

This is the problem that is being dealt with here.

Good point, thank you

All the test are filling up available extents. We should see extents left behind and ensure that they are in the correct state.

deadalnix · 2023-10-24T23:12:58Z

sdlib/d/gc/bin.d

+
+		auto size = getSizeFromClass(sc);
+		auto gotSlots =
+			arena.allocSmallBatch(&emap, size, buffer[0 .. wantSlots]);


The entry point should be thing bin's batchAllocate .

deadalnix · 2023-10-24T23:56:36Z

sdlib/d/gc/bin.d

+	auto expectedSlabs = 4;
+
+	import d.gc.sizeclass;
+	foreach (sc; 0 .. ClassCount.Small) {


On the other hand, checking all size class isn't really necessary. Checking a couple, with a couple of slot count could be useful.

deadalnix · 2023-11-06T21:31:28Z

As mentioned in #334 (comment) , the title/description is not useful in its current shape.

deadalnix · 2023-11-06T21:33:33Z

sdlib/d/gc/bin.d

-		if (slab is null) {
-			return null;
-		}
+		auto remainBuffer = buffer;


If you need to create a local variable to track thing, it'd be simple for it to be the number of items allocated.

You'd avoid a cast and various computations.

dsm9000 force-pushed the batch_alloc_from_bin branch 5 times, most recently from 9b27c78 to 6bd2c62 Compare October 24, 2023 04:15

dsm9000 marked this pull request as draft October 24, 2023 14:16

dsm9000 force-pushed the batch_alloc_from_bin branch from 6bd2c62 to dfd2f71 Compare October 24, 2023 14:55

dsm9000 marked this pull request as ready for review October 24, 2023 14:55

dsm9000 force-pushed the batch_alloc_from_bin branch 2 times, most recently from 88c246a to f4e8974 Compare October 24, 2023 18:09

dsm9000 marked this pull request as draft October 24, 2023 18:32

dsm9000 force-pushed the batch_alloc_from_bin branch from f4e8974 to c20c78c Compare October 24, 2023 18:56

dsm9000 marked this pull request as ready for review October 24, 2023 18:56

dsm9000 force-pushed the batch_alloc_from_bin branch from c20c78c to ca00271 Compare October 24, 2023 18:59

deadalnix reviewed Oct 24, 2023

View reviewed changes

dsm9000 force-pushed the batch_alloc_from_bin branch from ca00271 to 1d7e8c8 Compare October 25, 2023 03:44

Batch alloc from bin with behaviour patterned after jemalloc's

73f9a28

dsm9000 force-pushed the batch_alloc_from_bin branch from 1d7e8c8 to 73f9a28 Compare October 25, 2023 03:45

deadalnix reviewed Nov 6, 2023

View reviewed changes

deadalnix force-pushed the master branch from 07c29a7 to 72f7b01 Compare November 8, 2023 15:19

deadalnix force-pushed the master branch from ead67da to cb7b965 Compare November 16, 2023 12:37

deadalnix force-pushed the master branch 3 times, most recently from cf07b66 to 39f39ba Compare January 3, 2024 14:18

deadalnix force-pushed the master branch from 2e804ec to 416d7d7 Compare January 12, 2024 01:22

deadalnix force-pushed the master branch 2 times, most recently from 4acd1ab to e854d3d Compare June 17, 2024 21:01

deadalnix force-pushed the master branch from 2dd00d4 to 0a7c915 Compare July 24, 2024 23:58

deadalnix force-pushed the master branch from 003b3b5 to df865c3 Compare September 27, 2024 21:13

deadalnix force-pushed the master branch 3 times, most recently from 0896895 to 5a79292 Compare October 13, 2024 11:52

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Batch alloc from bin with behaviour patterned after jemalloc's #334

Batch alloc from bin with behaviour patterned after jemalloc's #334

dsm9000 commented Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

dsm9000 Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

dsm9000 Oct 24, 2023

deadalnix Oct 25, 2023

dsm9000 Oct 25, 2023 •

edited

Loading

deadalnix commented Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

dsm9000 Oct 24, 2023

deadalnix Nov 6, 2023

deadalnix Oct 24, 2023

deadalnix Oct 24, 2023

deadalnix commented Nov 6, 2023

deadalnix Nov 6, 2023

deadalnix Nov 6, 2023

		auto toFill = needFill - filled;
		auto count = toFill < nFree ? toFill : nFree;

Batch alloc from bin with behaviour patterned after jemalloc's #334

Are you sure you want to change the base?

Batch alloc from bin with behaviour patterned after jemalloc's #334

Conversation

dsm9000 commented Oct 24, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dsm9000 Oct 25, 2023 • edited Loading

Choose a reason for hiding this comment

deadalnix commented Oct 24, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

deadalnix commented Nov 6, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dsm9000 Oct 25, 2023 •

edited

Loading