cmd/go: limit total compile -c backend concurrency using a pool

Previously we limited the value we passed in to compile -c (which set
the number of SSA compile goroutines that run at one time) to 4. This CL
allows the -c value to go up to GOMAXPROCS, while limiting the total
number of backend SSA compile goroutines to still be less than the
previous worst case of 4*GOMAXPROCS (actually four times the value of
the -p flag, but the default is GOMAXPROCS). We do that by keeping a
pool of tokens to represent the total number of SSA compile goroutines
(with some buffer to allow us to run out of tokens and not exceed
4*GOMAXPROCS). Each time a compile requests a -c value, we'll hand out
half of the remaining tokens (with the number handed otu capped at
GOMAXPROCS) until we run out of tokens, in wich case we'll set -c to
one.

This leads to a speed up of 3-10% on the 16 core intel perf builder and
5-16% on the 88 core builder on the Sweet go-build benchmark.

Change-Id: Ib1ec843fee57f0fb8d36a507162317276a6a6964
Reviewed-on: https://go-review.googlesource.com/c/go/+/724142
Reviewed-by: Alan Donovan <adonovan@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Matloob <matloob@google.com>

src/cmd/go/internal/work/gc.go

@@ -15,6 +15,7 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+	"sync"
 
 	"cmd/go/internal/base"
 	"cmd/go/internal/cfg"
@@ -127,7 +128,9 @@ func (gcToolchain) gc(b *Builder, a *Action, archive string, importcfg, embedcfg
 		gcflags = append(gcflags, fuzzInstrumentFlags()...)
 	}
 	// Add -c=N to use concurrent backend compilation, if possible.
-	if c := gcBackendConcurrency(gcflags); c > 1 {
+	c, release := compilerConcurrency()
+	defer release()
+	if c > 1 {
 		defaultGcFlags = append(defaultGcFlags, fmt.Sprintf("-c=%d", c))
 	}
 
@@ -177,8 +180,9 @@ func (gcToolchain) gc(b *Builder, a *Action, archive string, importcfg, embedcfg
 	return ofile, output, err
 }
 
-// gcBackendConcurrency returns the backend compiler concurrency level for a package compilation.
+// compilerConcurrency returns the compiler concurrency level for a package compilation.
-func gcBackendConcurrency(gcflags []string) int {
+// The returned function must be called after the compile finishes.
+func compilerConcurrency() (int, func()) {
 	// First, check whether we can use -c at all for this compilation.
 	canDashC := concurrentGCBackendCompilationEnabledByDefault
 
@@ -199,7 +203,7 @@ func gcBackendConcurrency(gcflags []string) int {
 	}
 
 	if !canDashC {
-		return 1
+		return 1, func() {}
 	}
 
 	// Decide how many concurrent backend compilations to allow.
@@ -212,29 +216,60 @@ func gcBackendConcurrency(gcflags []string) int {
 	// of the overall compiler execution, so c==1 for much of the build.
 	// So don't worry too much about that interaction for now.
 	//
-	// However, in practice, setting c above 4 tends not to help very much.
+	// But to keep things reasonable, we maintain a cap on the total number of
-	// See the analysis in CL 41192.
+	// concurrent backend compiles. (If we gave each compile action the full GOMAXPROCS, we could
+	// potentially have GOMAXPROCS^2 running compile goroutines) In the past, we'd limit
+	// the number of concurrent backend compiles per process to 4, which would result in a worst-case number
+	// of backend compiles of 4*cfg.BuildP. Because some compile processes benefit from having
+	// a larger number of compiles, especially when the compile action is the only
+	// action running, we'll allow the max value to be larger, but ensure that the
+	// total number of backend compiles never exceeds that previous worst-case number.
+	// This is implemented using a pool of tokens that are given out. We'll set aside enough
+	// tokens to make sure we don't run out, and then give half of the remaining tokens (up to
+	// GOMAXPROCS) to each compile action that requests it.
 	//
-	// TODO(josharian): attempt to detect whether this particular compilation
+	// As a user, to limit parallelism, set GOMAXPROCS below numCPU; this may be useful
-	// is likely to be a bottleneck, e.g. when:
-	//   - it has no successor packages to compile (usually package main)
-	//   - all paths through the build graph pass through it
-	//   - critical path scheduling says it is high priority
-	// and in such a case, set c to runtime.GOMAXPROCS(0).
-	// By default this is the same as runtime.NumCPU.
-	// We do this now when p==1.
-	// To limit parallelism, set GOMAXPROCS below numCPU; this may be useful
 	// on a low-memory builder, or if a deterministic build order is required.
-	c := runtime.GOMAXPROCS(0)
 	if cfg.BuildP == 1 {
 		// No process parallelism, do not cap compiler parallelism.
-		return c
+		return maxCompilerConcurrency, func() {}
 	}
-	// Some process parallelism. Set c to min(4, maxprocs).
+
-	if c > 4 {
+	// Cap compiler parallelism using the pool.
-		c = 4
+	tokensMu.Lock()
+	defer tokensMu.Unlock()
+	concurrentProcesses++
+	// Set aside tokens so that we don't run out if we were running cfg.BuildP concurrent compiles.
+	// We'll set aside one token for each of the action goroutines that aren't currently running a compile.
+	setAside := cfg.BuildP - concurrentProcesses
+	availableTokens := tokens - setAside
+	// Grab half the remaining tokens: but with a floor of at least 1 token, and
+	// a ceiling of the max backend concurrency.
+	c := max(min(availableTokens/2, maxCompilerConcurrency), 1)
+	tokens -= c
+	// Successfully grabbed the tokens.
+	return c, func() {
+		tokensMu.Lock()
+		defer tokensMu.Unlock()
+		tokens += c
 	}
-	return c
+}
+
+var maxCompilerConcurrency = runtime.GOMAXPROCS(0) // max value we will use for -c
+
+var (
+	tokensMu            sync.Mutex
+	tokens              int // number of available tokens
+	concurrentProcesses int // number of currently running compiles
+)
+
+// initCompilerConcurrencyPool sets the number of tokens in the pool. It needs
+// to be run after init, so that it can use the value of cfg.BuildP.
+func initCompilerConcurrencyPool() {
+	// Size the pool so that the worst case total number of compiles is not more
+	// than what it was when we capped the concurrency to 4.
+	oldConcurrencyCap := min(4, maxCompilerConcurrency)
+	tokens = oldConcurrencyCap * cfg.BuildP
 }
 
 // trimpath returns the -trimpath argument to use

src/cmd/go/internal/work/init.go

@@ -60,6 +60,7 @@ func BuildInit(loaderstate *modload.State) {
 	modload.Init(loaderstate)
 	instrumentInit()
 	buildModeInit()
+	initCompilerConcurrencyPool()
 	cfgChangedEnv = makeCfgChangedEnv()
 
 	if err := fsys.Init(); err != nil {