The test result of oslat with task isolation patches

run-oslat-isol.sh

sudo tuned-adm profile latency-performance

OUTPUT=output-oslat-`uname -r`-`date +%s`

stress-ng --job stress-jobs &
STRESS=$!
sleep 5

# chisol and oslat here are compiled with task isolation userspace patches
# which can be found at https://people.redhat.com/~mtosatti/task-isol-v6-userspace-patches/
chisol -q vmstat_sync -I conf oslat -c 2 -b 400 -w memmove -m 16k -D 15m --json=${OUTPUT}

sudo kill -2 ${STRESS}

echo "Save the result to ${OUTPUT}"

run-oslat.sh

sudo tuned-adm profile latency-performance

OUTPUT=output-oslat-`uname -r`-`date +%s`

stress-ng --job stress-jobs &
STRESS=$!
sleep 5

# This is the original oslat compiled without task isolation userspace patch
oslat -c 2 -b 400 -w memmove -m 16k -D 15m --json=${OUTPUT}

sudo kill -2 ${STRESS}

echo "Save the result to ${OUTPUT}"

stress-jobs-sample

# This is stress-ng config used for test 4 (vm + mmap + fault)

taskset 0,1,3

fault 10
vm 2
vm-bytes 2G
mmap 2
mmap-bytes 2G
page-in

test-task-isolation-via-oslat.md

Testing task isolation via oslat

System information

Host

• 2 socket Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
• Ubuntu 20.04 with kernel 5.4.0-99-generic

Guest

• KVM with 4vCPU and 4GB RAM
• Debian 10 with kernel 5.15.18-rt28 with task isolation patches
• Command line: isolcpus=domain,nohz,2 nohz_full=2 irqaffinity=0,1,3
• oslat from rt-tests v2.30 with Marcelo's patch for oslat
• stress-ng v0.09.50

Each test spawns system workload on the CPUs which is not isolated, then runs a 15-min oslat with 16 KB memmove workload on the isolated CPU.

To compare the latency performance with and without task isolation, we perform each test on 2 different kernel configs. The one marked as TASKISOL implies that the kernel turns on CONFIG_TASK_ISOLATION.

Test 1: oslat on isolated CPU and timer workload on other CPUs

We create 20 timer workers in this test.

	PREEMPT_RT + NOHZ_FULL	TASKISOL + PREEMPT_RT + NOHZ_FULL
min (us)	2	2
avg (us)	2.007391	2.012791
max (us)	8391	30774
duration(sec)	899.182	899.183

Test 2: oslat on isolated CPU and major / minor fault workload on other CPUs

Since the patchset aims to eliminate the effect of vmstats, we create 20 fault workers as the system workload in this test.

	PREEMPT_RT + NOHZ_FULL	TASKISOL + PREEMPT_RT + NOHZ_FULL
min (us)	2	2
avg (us)	2.011208	2.015257
max (us)	5019	27741
duration(sec)	899.181	901.229

Test 3: oslat on isolated CPU and VM / mmap stress on other CPUs

To stress the system with large VM stuff, we use 2 vm and 2 mmap workers. Each worker uses 2GB to raise OOM frequently.

	PREEMPT_RT + NOHZ_FULL	TASKISOL + PREEMPT_RT + NOHZ_FULL
min (us)	2	2
avg (us)	2.016188	2.02046
max (us)	10153	12648
duration(sec)	899.182	901.229

Test 4: oslat on isolated CPU and VM / mmap / fault workload on other CPUs

This test uses a mixed workload comprised of 10 fault, 2 vm and 2 mmap workers.

	PREEMPT_RT + NOHZ_FULL	TASKISOL + PREEMPT_RT + NOHZ_FULL
min (us)	2	2
avg (us)	2.014711	2.013344
max (us)	12908	12707
duration(sec)	899.183	899.184