CRC32 for short atomic strings

[Anton Bikineev]

Hi all,

I wrote a short doc on how we can speed up insertion into the AtomicStringTable,

which is in particular hot on VanillaJS Speedometer3 stories. The idea is:

 - use crc32 for short strings (one-word and two-words) for hashing, which is

   faster for short strings than rapidhash (see the doc),

 - store short strings in separate tables and cache them inlined in the backing,

   to avoid indirection on comparison.

On Linux I see a 33% speedup for AtomicStringTable::Add() when running the S3

TodoMVC-JavaScript.* stories. On M1 the S3 total score improvement is 0.3%.

[Kentaro Hara]

LGTM. This is an internal optimization scoped to the hashing algorithm and I have no concern.

--
You received this message because you are subscribed to the Google Groups "platform-architecture-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to platform-architect...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/platform-architecture-dev/CABH6udYDACF2gpB6X43xYngBjt107FpDJkDVqW%2BsrTOGji%3Dsiw%40mail.gmail.com.

--

Kentaro Hara, Tokyo

[Dave Tapuska]

Do we know if this is limited to a set of strings? (I did produce a histogram of strings before)

When I last looked at speedometer2, we were inserting and removing data-id from the atomic string table frequently... See https://chromium-review.googlesource.com/c/chromium/src/+/5784168 which I never landed.

This change alone had a 0.3% improvement on speedometer2 at the time. I understand this is speedometer3 though, so it might be different.

dave.

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/platform-architecture-dev/CABg10jxgsRmaS8RQw0iMikQ%2B7t%2BWtsRW2t%3D5%3Dds6EYBTRakXKA%40mail.gmail.com.

[Ian Kilpatrick]

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/platform-architecture-dev/CAHgVhZXv7D0D_s4kvhE3kpj-yfj-kpLC47JZvga%2BDuRT9aot0A%40mail.gmail.com.

[Anton Bikineev]

Do we know if this is limited to a set of strings? (I did produce a histogram of strings before)

The short strings on S3 are pretty much random, I just pasted them here.

[Anton Bikineev]

When I last looked at speedometer2, we were inserting and removing data-id from the atomic string table frequently... See https://chromium-review.googlesource.com/c/chromium/src/+/5784168 which I never landed. 

This change alone had a 0.3% improvement on speedometer2 at the time. I understand this is speedometer3 though, so it might be different.

The idea of crc32 with extra tables with inlined storage is orthogonal in that it won't speed up string creation/destruction. It's meant to make the lookups in the table faster. So, I think your CL may still matter for S3.

[Michael Lippautz]

LGTM. While this is some complexity it's well encapsulated and doesn't affect anything outside of the AtomicStringTable.

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/platform-architecture-dev/CABH6udZ9UHOZHJkgYN7vf%3Dv%2Bkam7KCjMCX4ikXQMkgSdPO4XQw%40mail.gmail.com.