Commit 19377d66 authored by Tomasz Zawadzki's avatar Tomasz Zawadzki
Browse files

test/scheduler: measure ticks since last check 



get_thread_stats() gathers busy/idle from start of
the application, directly from RPC.
For some tests it is not required to measure state
from app start, but only changes during test.

idle.sh assumes that second check would make all
threads idle. Yet if initialization takes long enough,
app_thread will not be idle longer than it was busy by then.
This became apparent when adding ASAN to all jobs,
which considerably increased start up time.

To fix that, only measure ticks between consecutive RPC calls.
Added get_thread_stats_current() that returns busy/idle
difference, since last time it was called.

It allows to remove the workaround for first sample,
per removed comment.

While here:
- added print with load/busy/idle for each thread
- fixed unlikely condition for idle[thread] == busy[thread]

Signed-off-by: default avatarTomasz Zawadzki <tomasz.zawadzki@intel.com>
Change-Id: Iaacce31aa7d4bcc22a51dbd7e71a4c7a4d274862
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/20884


Reviewed-by: default avatarMichal Berger <michal.berger@intel.com>
Reviewed-by: default avatarBen Walker <ben@nvidia.com>
Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarJim Harris <jim.harris@samsung.com>
parent 8cd4cc63

test/scheduler/common.sh

+16 −0
Original line number Diff line number Diff line
@@ -406,6 +406,22 @@ exec_under_dynamic_scheduler() { 
	"$rootdir/scripts/rpc.py" framework_start_init

}



# Gather busy/idle stats since this function was last called

get_thread_stats_current() {

	xtrace_disable



	local total_busy total_idle



	_get_thread_stats total_busy total_idle



	for thread in "${!thread_map[@]}"; do

		: $((busy[thread] = total_busy[thread] - past_busy[thread], past_busy[thread] = total_busy[thread]))

		: $((idle[thread] = total_idle[thread] - past_idle[thread], past_idle[thread] = total_idle[thread]))

	done

	xtrace_restore

}



# Gather busy/idle stats since application start

get_thread_stats() {

	xtrace_disable

	_get_thread_stats busy idle

test/scheduler/idle.sh

+10 −7
Original line number Diff line number Diff line
@@ -13,17 +13,20 @@ source "$testdir/common.sh" 
trap 'killprocess "$spdk_pid"' EXIT



thread_stats() {

	local thread

	local thread load

	busy_threads=0



	get_thread_stats

	get_thread_stats_current



	# Simply verify if threads stay idle

	for thread in "${!thread_map[@]}"; do

		printf '[load:%3u%%, idle:%10u, busy:%10u] ' \

			$((busy[thread] * 100 / (busy[thread] + idle[thread]))) \

			"${idle[thread]}" "${busy[thread]}"

		if ((idle[thread] < busy[thread])); then

			printf 'Waiting for %s to become idle\n' "${thread_map[thread]}"

			((++busy_threads))

		elif ((idle[thread] > busy[thread])); then

		else

			printf '%s is idle\n' "${thread_map[thread]}"

		fi

	done
@@ -41,6 +44,9 @@ idle() { 
	# - all threads are assigned to main lcore

	# - threads are not being moved between lcores



	# Get first set of stats, to exclude initialization from the busy/idle

	get_thread_stats_current



	xtrace_disable

	while ((samples++ < 5)); do

		cpumask=0
@@ -56,10 +62,7 @@ idle() { 


		thread_stats



		# Allow app_thread is busy for the first sample. Because on some system the dpdk_governor

		# initiation process on app_thread is time consuming. This may make the busy time greater

		# than idle time which causes the test to fail.

		((samples == 1 && busy_threads <= 1 || samples > 1 && busy_threads == 0))

		((busy_threads == 0))

	done



	xtrace_restore