alts: Fix buffer alignment with 16KB records (#8791)

gRPC Go receives ALTS records of max 16KB under high load (see
https://github.com/grpc/grpc-go/pull/8512#issuecomment-3193280949 for
details). When the ALTS conn has a partial encrypted frame in its
buffer, it attempts to copy the frame to the beginning of the buffer to
read the remainder.

https://github.com/grpc/grpc-go/blob/40466769682557e7179b8c74ba3820cc78d49b4b/credentials/alts/internal/conn/record.go#L151-L159

When using a buffer of exactly 32KiB, almost the entire second frame of
16KiB is stored, but not the full frame. As a result, a large copy of
~16KiB is performed.

## Solution
This PR increases the read buffer length by 512 bytes to ensure two
entire 16KiB frames can be stored. This ensures that usually, only ~512
bytes needs to be moved to the front.

## Benchmark
In a GCS directpath benchmark downloading files in a loop, the time
spent on memory copies in the ALTS code is eliminated, saving ~3.7% of
CPU time.

### Before
<img width="1532" height="480" alt="image"
src="https://github.com/user-attachments/assets/c2051b7f-e828-44d6-a82a-0d444cace2df"
/>

### After
<img width="1550" height="690" alt="image"
src="https://github.com/user-attachments/assets/5715c5b8-b163-474c-81a6-5bcc6808d2ff"
/>

RELEASE NOTES:
* credentials/alts: optimize read buffer alignment to reduce copies.

diff --git a/credentials/alts/internal/conn/record.go b/credentials/alts/internal/conn/record.go
index f9d2646d4b22f5e94a835c61497624cce789f88d..329e3691f0d55bb432447390d18160282575de02 100644 (file)
--- a/credentials/alts/internal/conn/record.go
+++ b/credentials/alts/internal/conn/record.go
@@ -68,7 +68,9 @@ const (
        // altsRecordDefaultLength.
        altsWriteBufferMaxSize = 512 * 1024 // 512KiB
        // The initial buffer used to read from the network.
-       altsReadBufferInitialSize = 32 * 1024 // 32KiB
+       // It includes an additional 512 Bytes to hold two 16KiB records plus
+       // small framing overheads.
+       altsReadBufferInitialSize = 32*1024 + 512 // 32.5KiB
 )
 
 var (