ivampiresp/Spdk
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

test/scheduler: Fallback to checking user mode load if idle check fails

Browse Source 
During tests, the main portion of the app's time is spent in user
mode when polling is in effect. Currently, the idleness of given cpu
is determined in comparision to all the possible states the cpu can
be in. This makes it easier for third party processes to impact the
idleness of the cpu (as it may be lowered via different kinds of
loads).

For instance, if target cpu suddenly sees spike in a system
load (even up to 100%) it should not be relevant for the test where
it's known that the app's polling utilized mainly user mode prior
switching the scheduler.

With that in mind, if the general idle check fails, lookup the
raw samples matching the user mode and if the load is relatively
low (<= 15%) consider this as a pass.

Signed-off-by: Michal Berger <michallinuxstuff@gmail.com>
Change-Id: I4ab260d8bcf20a69f2f0be10f0fd7be577682be3
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12909
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
This commit is contained in:
Michal Berger 2022-06-06 01:34:53 +02:00 committed by Tomasz Zawadzki
parent 7d080339ec
commit 9b8939660a
1 changed files with 48 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
test/scheduler/common.sh
View File

@ -460,7 +460,7 @@ get_cpu_time() {
# 8 - guest (Time spent running a virtual CPU)
# 9 - guest_nice (Time spent running a niced guest)
local -A cpu_time_map
local -gA cpu_time_map
cpu_time_map["user"]=0
cpu_time_map["nice"]=1
cpu_time_map["system"]=2
@ -477,6 +477,7 @@ get_cpu_time() {
unset -v ${!old_stat_@}
unset -v ${!avg_stat@}
unset -v ${!avg_load@}
unset -v ${!raw_samples@}
cpu_time=${cpu_time_map["$cpu_time"]}
interval=$((interval <= 0 ? 1 : interval))
@ -486,6 +487,8 @@ get_cpu_time() {
for cpu in "${cpus[@]}"; do
local -n old_stats=old_stats_$cpu
local -n avg_load=avg_load_$cpu
local -n raw_samples=raw_samples_$cpu
sample_stats=() total_sample=0
stats=($(get_cpu_stat "$cpu" all))
@ -501,6 +504,9 @@ get_cpu_time() {
done
for stat in "${!stats[@]}"; do
local -n avg_stat=stat_${stat}_${cpu}
local -n raw_samples_ref=raw_samples_${stat}_${cpu}
raw_samples[stat]="raw_samples_${stat}_${cpu}[@]"
raw_samples_ref+=("${stats[stat]}")
avg_stat+=($((sample_stats[stat] * 100 / (total_sample == 0 ? 1 : total_sample))))
done
old_stats=("${stats[@]}")
@ -537,18 +543,26 @@ collect_cpu_idle() {
get_cpu_time "$time" idle "${cpus_to_collect[@]}"
local user_load
for cpu in "${cpus_to_collect[@]}"; do
samples=(${cpu_times[cpu]})
printf '* cpu%u idle samples: %s (avg: %u%%)\n' \
"$cpu" "${samples[*]}" "${avg_cpu_time[cpu]}"
# Cores with polling reactors have 0% idle time,
# while the ones in interrupt mode won't have 100% idle.
# Work can be potentially be scheduled to the core by kernel,
# to prevent that affecting tests set reasonably high idle limit.
# Consider last sample
# During the tests, polling reactors spend the major portion
# of their cpu time in user mode. With that in mind, if the
# general check for cpus's idleness fails, check what portion
# of the cpu load falls into user mode. For the idle check
# use the last sample. For the cpu load, compare user's raw
# samples in SC_CLK_TCK context for a more detailed view.
user_load=$(cpu_usage_clk_tck "$cpu" user)
if ((samples[-1] >= 70)); then
printf '* cpu%u is idle\n' "$cpu"
is_idle[cpu]=1
elif ((user_load <= 15)); then
printf '* cpu%u not fully idle, but user load is low so passing\n' "$cpu"
is_idle[cpu]=1
else
printf '* cpu%u is not idle\n' "$cpu"
is_idle[cpu]=0
@ -556,6 +570,36 @@ collect_cpu_idle() {
done
}
cpu_usage_clk_tck() {
local cpu=$1 time=${2:-all}
local user nice system usage clk_delta
# We should be called in get_cpu_time()'s environment.
[[ -v raw_samples_$cpu ]] || return 1
local -n raw_samples=raw_samples_$cpu
user=("${!raw_samples[cpu_time_map["user"]]}")
nice=("${!raw_samples[cpu_time_map["nice"]]}")
system=("${!raw_samples[cpu_time_map["system"]]}")
# Construct delta based on last two samples of a given time.
case "$time" in
user | all) ((clk_delta += (user[-1] - user[-2]))) ;;&
nice | all) ((clk_delta += (nice[-1] - nice[-2]))) ;;&
system | all) ((clk_delta += (system[-1] - system[-2]))) ;;
*) ;;
esac
# We assume 1s between each sample. See get_cpu_time().
usage=$((100 * clk_delta / $(getconf CLK_TCK)))
usage=$((usage > 100 ? 100 : usage))
printf '%u' "$usage"
printf '* cpu%u %s usage: %u\n' "$cpu" "$time" "$usage" >&2
printf '* cpu%u user samples: %s\n' "$cpu" "${user[*]}" >&2
printf '* cpu%u nice samples: %s\n' "$cpu" "${nice[*]}" >&2
printf '* cpu%u system samples: %s\n' "$cpu" "${system[*]}" >&2
}
update_thread_cpus_map() {
local cpu
local -g thread_cpus=()