internal/trace: improve gc-stress test

The gc-stress test is useful for trying to exercise GC-related trace
events by producing a lot of them in many different situations.
Unfortunately this test is flaky, because allocating in a loop can
easily out-run the GC when it's trying to preempt the allocating
goroutine.

It's been a long standing problem that a program that allocates in a
loop can outrun a GC. The problem isn't the GC persay, it's consistently
correlated with a high STW time (likely a high 'stopping' time, not a
'stopped' time), suggesting that in the window of time when the garbage
collector is trying to stop all goroutines, they continue to allocate.

This should probably be fixed in general, but for now, let's focus on
this flaky test.

This CL changes the gc-stress test to (1) set a memory limit and (2) do
more work in between allocations. (2) is really what makes things less
flaky, but (2) unfortunately also means the GC is less exercised. That's
where (1) comes in. By setting a low memory limit, we increase GC
activity (in particular, assist activity). The memory limit also helps
prevent the heap from totally blowing up due to the heap goal inflating
from floating garbage, but it's not perfect.

After this change, under stress2, this test exceeds a heap size of 500
MiB only 1 in 5000 runs on my 64-vCPU VM. Before this change, it got
that big about 1/4th of the time.

Fixes #74052.

Change-Id: I49233c914c8b65b1d593d3953891fddda6685aec
Reviewed-on: https://go-review.googlesource.com/c/go/+/683515
Reviewed-by: Carlos Amedee <carlos@golang.org>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>

diff --git a/src/internal/trace/testdata/testprog/gc-stress.go b/src/internal/trace/testdata/testprog/gc-stress.go
index 7979234c40a5714e3b614a219cab9cc058732f9a..74b63606d5e6199f53438ea67676a0dd674ee012 100644 (file)
--- a/src/internal/trace/testdata/testprog/gc-stress.go
+++ b/src/internal/trace/testdata/testprog/gc-stress.go
@@ -13,6 +13,7 @@ import (
        "log"
        "os"
        "runtime"
+       "runtime/debug"
        "runtime/trace"
        "time"
 )
@@ -36,11 +37,25 @@ func makeTree(depth int) *node {
        }
 }
 
+func initTree(n *node) {
+       if n == nil {
+               return
+       }
+       for i := range n.data {
+               n.data[i] = 'a'
+       }
+       for i := range n.children {
+               initTree(n.children[i])
+       }
+}
+
 var trees [16]*node
 var ballast *[16]*[1024]*node
-var sink [][]byte
+var sink []*node
 
 func main() {
+       debug.SetMemoryLimit(50 << 20)
+
        for i := range trees {
                trees[i] = makeTree(6)
        }
@@ -55,13 +70,17 @@ func main() {
        }
 
        procs := runtime.GOMAXPROCS(-1)
-       sink = make([][]byte, procs)
+       sink = make([]*node, procs)
 
        for i := 0; i < procs; i++ {
                i := i
                go func() {
                        for {
-                               sink[i] = make([]byte, 4<<10)
+                               sink[i] = makeTree(3)
+                               for range 5 {
+                                       initTree(sink[i])
+                                       runtime.Gosched()
+                               }
                        }
                }()
        }