Judge loss by arrival order and a deadline, one tx ring per stream

tx.go

@@ -1,6 +1,8 @@
 package main
 
 import (
+	"fmt"
+	"runtime"
 	"sync/atomic"
 
 	"golang.org/x/sys/unix"
@@ -19,6 +21,7 @@ type txStats struct {
 type txWorker struct {
 	fd      int
 	stream  uint16
+	cpu     int
 	spec    *frameSpec
 	batch   int
 	stats   *txStats
@@ -26,6 +29,16 @@ type txWorker struct {
 }
 
 func (w *txWorker) run(done *atomic.Bool) {
+	// XPS keys the hardware ring on the sending cpu, so the feeder lives on
+	// its own cpu for the stream's own ring: wire order within the stream is
+	// ring order, and one feeder stalling leaves the other stream's ring
+	// draining.
+	runtime.LockOSThread()
+	var set unix.CPUSet
+	set.Set(w.cpu)
+	if err := unix.SchedSetaffinity(0, &set); err != nil {
+		panic(fmt.Sprintf("pinning tx stream %d to cpu %d: %v", w.stream, w.cpu, err))
+	}
 	// The batch is not a multiple of the pattern count, so the pairing of pattern
 	// to size rotates every pass rather than settling into a fixed one.
 	bufs := make([][]byte, w.batch)