cmd/api: reduce parallel 'go list' invocations to a constant

'go list' has its own internal parallelism, so invoking in in parallel
can produce up to quadratic peak memory usage.

Running 'go list' is also very I/O-intensive, so the higher
parallelism does substantially improve latency; unfortunately, we lack
a good way to balance latency against memory footprint, so we need to
sacrifice some latency for reliability.

Fixes #49957.

Change-Id: Ib53990b46acf4cc67a9141644d97282964d6442d
Reviewed-on: https://go-review.googlesource.com/c/go/+/380994
Trust: Bryan Mills <bcmills@google.com>
Run-TryBot: Bryan Mills <bcmills@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org>
This commit is contained in:
Bryan C. Mills 2022-01-26 12:45:12 -05:00 committed by Bryan Mills
parent 719e9894f9
commit 946a8669d9
1 changed files with 5 additions and 2 deletions

View File

@ -459,8 +459,11 @@ type listImports struct {
var listCache sync.Map // map[string]listImports, keyed by contextName
// listSem is a semaphore restricting concurrent invocations of 'go list'.
var listSem = make(chan semToken, ((runtime.GOMAXPROCS(0)-1)/2)+1)
// listSem is a semaphore restricting concurrent invocations of 'go list'. 'go
// list' has its own internal concurrency, so we use a hard-coded constant (to
// allow the I/O-intensive phases of 'go list' to overlap) instead of scaling
// all the way up to GOMAXPROCS.
var listSem = make(chan semToken, 2)
type semToken struct{}