Tests showed, that under certain conditions, the summary number of jiffies
spent on softirq/idle, which are counted by system statistics can be even
below 10% of expected value, resulting in false load presentation.
The issue was observed on the quad-core Marvell Armada 8k SoC, whose two
10G ports were bound into L2 bridge. Load was controlled by bidirectional
UDP traffic, produced by a packet generator. Under such condition,
the dominant load is softirq. With 100% single CPU occupation or without
any activity (all CPUs 100% idle), total number of jiffies is 10000 (2500
per each core) in 10s interval. Also with other kind of load this was
true.
However below a saturation threshold it was observed, that with CPU which
was occupied almost by softirqs only, the statistic were awkward. See
the mpstat output:
CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
all 0.00 0.00 0.13 0.00 0.00 0.55 0.00 0.00 0.00 99.32
0 0.00 0.00 0.00 0.00 0.00 23.08 0.00 0.00 0.00 76.92
1 0.00 0.00 0.40 0.00 0.00 0.00 0.00 0.00 0.00 99.60
2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00
3 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00
Above would mean basically no total load, debug CPU0 occupied in 25%.
Raw statistics, printed every 10s from /proc/stat unveiled a root
cause - summary idle/softirq jiffies on loaded CPU were below 200,
i.e. over 90% samples lost. All problems were gone after enabling
fine granulity IRQ time accounting.
This patch fixes possible wrong statistics processing by enabling
CONFIG_IRQ_TIME_ACCOUNTING for arm64 platfroms, which is by
default done on other architectures, e.g. x86 and arm. Tests
showed no noticeable performance penalty, nor stability impact.
Signed-off-by: Marcin Wojtas <mw@semihalf.com>
Signed-off-by: Gregory CLEMENT <gregory.clement@free-electrons.com>
CONFIG_AUDIT=y
CONFIG_NO_HZ_IDLE=y
CONFIG_HIGH_RES_TIMERS=y
+CONFIG_IRQ_TIME_ACCOUNTING=y
CONFIG_BSD_PROCESS_ACCT=y
CONFIG_BSD_PROCESS_ACCT_V3=y
CONFIG_TASKSTATS=y