diff --git a/src/runtime/os_windows.go b/src/runtime/os_windows.go
index a62e941..9dd140c 100644
--- a/src/runtime/os_windows.go
+++ b/src/runtime/os_windows.go
@@ -21,6 +21,7 @@
 //go:cgo_import_dynamic runtime._CreateIoCompletionPort CreateIoCompletionPort%4 "kernel32.dll"
 //go:cgo_import_dynamic runtime._CreateThread CreateThread%6 "kernel32.dll"
 //go:cgo_import_dynamic runtime._CreateWaitableTimerA CreateWaitableTimerA%3 "kernel32.dll"
+//go:cgo_import_dynamic runtime._CreateWaitableTimerExW CreateWaitableTimerExW%4 "kernel32.dll"
 //go:cgo_import_dynamic runtime._DuplicateHandle DuplicateHandle%7 "kernel32.dll"
 //go:cgo_import_dynamic runtime._ExitProcess ExitProcess%1 "kernel32.dll"
 //go:cgo_import_dynamic runtime._FreeEnvironmentStringsW FreeEnvironmentStringsW%1 "kernel32.dll"
@@ -68,6 +69,7 @@
 	_CreateIoCompletionPort,
 	_CreateThread,
 	_CreateWaitableTimerA,
+	_CreateWaitableTimerExW,
 	_DuplicateHandle,
 	_ExitProcess,
 	_FreeEnvironmentStringsW,
@@ -151,6 +153,8 @@
 	waitsema   uintptr // semaphore for parking on locks
 	resumesema uintptr // semaphore to indicate suspend/resume
 
+	highResTimer uintptr // high resolution timer handle used in usleep
+
 	// preemptExtLock synchronizes preemptM with entry/exit from
 	// external C code.
 	//
@@ -402,11 +406,21 @@
 // osRelax is called by the scheduler when transitioning to and from
 // all Ps being idle.
 //
-// On Windows, it adjusts the system-wide timer resolution. Go needs a
+// Some versions of Windows have high resolution timer. For those
+// versions osRelax is noop.
+// For Windows versions without high resolution timer, osRelax
+// adjusts the system-wide timer resolution. Go needs a
 // high resolution timer while running and there's little extra cost
 // if we're already using the CPU, but if all Ps are idle there's no
 // need to consume extra power to drive the high-res timer.
 func osRelax(relax bool) uint32 {
+	if haveHighResTimer {
+		// If the high resolution timer is available, the runtime uses the timer
+		// to sleep for short durations. This means there's no need to adjust
+		// the global clock frequency.
+		return 0
+	}
+
 	if relax {
 		return uint32(stdcall1(_timeEndPeriod, 1))
 	} else {
@@ -414,6 +428,42 @@
 	}
 }
 
+// haveHighResTimer indicates that the CreateWaitableTimerEx
+// CREATE_WAITABLE_TIMER_HIGH_RESOLUTION flag is available.
+var haveHighResTimer = false
+
+// createHighResTimer calls CreateWaitableTimerEx with
+// CREATE_WAITABLE_TIMER_HIGH_RESOLUTION flag to create high
+// resolution timer. createHighResTimer returns new timer
+// handle or 0, if CreateWaitableTimerEx failed.
+func createHighResTimer() uintptr {
+	const (
+		// As per @jstarks, see
+		// https://github.com/golang/go/issues/8687#issuecomment-656259353
+		_CREATE_WAITABLE_TIMER_HIGH_RESOLUTION = 0x00000002
+
+		_SYNCHRONIZE        = 0x00100000
+		_TIMER_QUERY_STATE  = 0x0001
+		_TIMER_MODIFY_STATE = 0x0002
+	)
+	return stdcall4(_CreateWaitableTimerExW, 0, 0,
+		_CREATE_WAITABLE_TIMER_HIGH_RESOLUTION,
+		_SYNCHRONIZE|_TIMER_QUERY_STATE|_TIMER_MODIFY_STATE)
+}
+
+func initHighResTimer() {
+	if GOARCH == "arm" {
+		// TODO: Not yet implemented.
+		return
+	}
+	h := createHighResTimer()
+	if h != 0 {
+		haveHighResTimer = true
+		usleep2Addr = unsafe.Pointer(funcPC(usleep2HighRes))
+		stdcall1(_CloseHandle, h)
+	}
+}
+
 func osinit() {
 	asmstdcallAddr = unsafe.Pointer(funcPC(asmstdcall))
 	usleep2Addr = unsafe.Pointer(funcPC(usleep2))
@@ -429,6 +479,7 @@
 
 	stdcall2(_SetConsoleCtrlHandler, funcPC(ctrlhandler), 1)
 
+	initHighResTimer()
 	timeBeginPeriodRetValue = osRelax(false)
 
 	ncpu = getproccount()
@@ -844,9 +895,20 @@
 	var thandle uintptr
 	stdcall7(_DuplicateHandle, currentProcess, currentThread, currentProcess, uintptr(unsafe.Pointer(&thandle)), 0, 0, _DUPLICATE_SAME_ACCESS)
 
+	// Configure usleep timer, if possible.
+	var timer uintptr
+	if haveHighResTimer {
+		timer = createHighResTimer()
+		if timer == 0 {
+			print("runtime: CreateWaitableTimerEx failed; errno=", getlasterror(), "\n")
+			throw("CreateWaitableTimerEx when creating timer failed")
+		}
+	}
+
 	mp := getg().m
 	lock(&mp.threadLock)
 	mp.thread = thandle
+	mp.highResTimer = timer
 	unlock(&mp.threadLock)
 
 	// Query the true stack base from the OS. Currently we're
@@ -884,6 +946,10 @@
 	lock(&mp.threadLock)
 	stdcall1(_CloseHandle, mp.thread)
 	mp.thread = 0
+	if mp.highResTimer != 0 {
+		stdcall1(_CloseHandle, mp.highResTimer)
+		mp.highResTimer = 0
+	}
 	unlock(&mp.threadLock)
 }
 
@@ -976,9 +1042,12 @@
 	return stdcall(fn)
 }
 
-// in sys_windows_386.s and sys_windows_amd64.s
+// In sys_windows_386.s and sys_windows_amd64.s.
 func onosstack(fn unsafe.Pointer, arg uint32)
+
+// These are not callable functions. They should only be called via onosstack.
 func usleep2(usec uint32)
+func usleep2HighRes(usec uint32)
 func switchtothread()
 
 var usleep2Addr unsafe.Pointer
diff --git a/src/runtime/sys_windows_386.s b/src/runtime/sys_windows_386.s
index 9e1f409..4ac1527 100644
--- a/src/runtime/sys_windows_386.s
+++ b/src/runtime/sys_windows_386.s
@@ -428,6 +428,42 @@
 	MOVL	BP, SP
 	RET
 
+// Runs on OS stack. duration (in 100ns units) is in BX.
+TEXT runtime·usleep2HighRes(SB),NOSPLIT,$36
+	// Want negative 100ns units.
+	NEGL	BX
+	MOVL	$-1, hi-4(SP)
+	MOVL	BX, lo-8(SP)
+
+	get_tls(CX)
+	MOVL	g(CX), CX
+	MOVL	g_m(CX), CX
+	MOVL	(m_mOS+mOS_highResTimer)(CX), CX
+	MOVL	CX, saved_timer-12(SP)
+
+	MOVL	$0, fResume-16(SP)
+	MOVL	$0, lpArgToCompletionRoutine-20(SP)
+	MOVL	$0, pfnCompletionRoutine-24(SP)
+	MOVL	$0, lPeriod-28(SP)
+	LEAL	lo-8(SP), BX
+	MOVL	BX, lpDueTime-32(SP)
+	MOVL	CX, hTimer-36(SP)
+	MOVL	SP, BP
+	MOVL	runtime·_SetWaitableTimer(SB), AX
+	CALL	AX
+	MOVL	BP, SP
+
+	MOVL	$0, ptime-28(SP)
+	MOVL	$0, alertable-32(SP)
+	MOVL	saved_timer-12(SP), CX
+	MOVL	CX, handle-36(SP)
+	MOVL	SP, BP
+	MOVL	runtime·_NtWaitForSingleObject(SB), AX
+	CALL	AX
+	MOVL	BP, SP
+
+	RET
+
 // Runs on OS stack.
 TEXT runtime·switchtothread(SB),NOSPLIT,$0
 	MOVL	SP, BP
diff --git a/src/runtime/sys_windows_amd64.s b/src/runtime/sys_windows_amd64.s
index 6c8eecd..8475425 100644
--- a/src/runtime/sys_windows_amd64.s
+++ b/src/runtime/sys_windows_amd64.s
@@ -457,6 +457,38 @@
 	MOVQ	40(SP), SP
 	RET
 
+// Runs on OS stack. duration (in 100ns units) is in BX.
+TEXT runtime·usleep2HighRes(SB),NOSPLIT|NOFRAME,$72
+	MOVQ	SP, AX
+	ANDQ	$~15, SP	// alignment as per Windows requirement
+	MOVQ	AX, 64(SP)
+
+	get_tls(CX)
+	MOVQ	g(CX), CX
+	MOVQ	g_m(CX), CX
+	MOVQ	(m_mOS+mOS_highResTimer)(CX), CX	// hTimer
+	MOVQ	CX, 48(SP)				// save hTimer for later
+	// Want negative 100ns units.
+	NEGQ	BX
+	LEAQ	56(SP), DX				// lpDueTime
+	MOVQ	BX, (DX)
+	MOVQ	$0, R8					// lPeriod
+	MOVQ	$0, R9					// pfnCompletionRoutine
+	MOVQ	$0, AX
+	MOVQ	AX, 32(SP)				// lpArgToCompletionRoutine
+	MOVQ	AX, 40(SP)				// fResume
+	MOVQ	runtime·_SetWaitableTimer(SB), AX
+	CALL	AX
+
+	MOVQ	48(SP), CX				// handle
+	MOVQ	$0, DX					// alertable
+	MOVQ	$0, R8					// ptime
+	MOVQ	runtime·_NtWaitForSingleObject(SB), AX
+	CALL	AX
+
+	MOVQ	64(SP), SP
+	RET
+
 // Runs on OS stack.
 TEXT runtime·switchtothread(SB),NOSPLIT|NOFRAME,$0
 	MOVQ	SP, AX
diff --git a/src/runtime/sys_windows_arm.s b/src/runtime/sys_windows_arm.s
index 256b5ff..57415e1 100644
--- a/src/runtime/sys_windows_arm.s
+++ b/src/runtime/sys_windows_arm.s
@@ -468,6 +468,11 @@
 	MOVW	R4, R13			// Restore SP
 	MOVM.IA.W (R13), [R4, R15]	// pop {R4, pc}
 
+// Runs on OS stack. Duration (in 100ns units) is in R0.
+// TODO: neeeds to be implemented properly.
+TEXT runtime·usleep2HighRes(SB),NOSPLIT|NOFRAME,$0
+	B	runtime·abort(SB)
+
 // Runs on OS stack.
 TEXT runtime·switchtothread(SB),NOSPLIT|NOFRAME,$0
 	MOVM.DB.W [R4, R14], (R13)  	// push {R4, lr}