[PATCH 0/2] Exposing nice CPU usage to userspace

[PATCH 0/2] Exposing nice CPU usage to userspace

[Joshua]

From: Joshua Hahn <joshua.hahn6@gmail.com>

Niced CPU usage is a metric reported in host-level /proc/stat, but is
not reported in cgroup-level statistics in cpu.stat. However, when a
host contains multiple tasks across different workloads, it becomes
difficult to gauage how much of the task is being spent on niced
processes based on /proc/stat alone, since host-level metrics do not
provide this cgroup-level granularity.

Exposing this metric will allow load balancers to correctly probe the
niced CPU metric for each workload, and make more informed decisions
when directing higher priority tasks.

Joshua Hahn (2):
  Tracking cgroup-level niced CPU time
  Selftests for niced CPU statistics

 include/linux/cgroup-defs.h               |  1 +
 kernel/cgroup/rstat.c                     | 16 ++++-
 tools/testing/selftests/cgroup/test_cpu.c | 72 +++++++++++++++++++++++
 3 files changed, 86 insertions(+), 3 deletions(-)

-- 
2.43.5

[PATCH 1/2] Tracking cgroup-level niced CPU time

[Joshua]

From: Joshua Hahn <joshua.hahn6@gmail.com>

Cgroup-level CPU statistics currently include time spent on
user/system processes, but do not include niced CPU time (despite
already being tracked). This patch exposes niced CPU time to the
userspace, allowing users to get a better understanding of their
hardware limits and can facilitate better load-balancing.

A new field 'ntime' is added to struct cgroup_base_stat as opposed to
struct task_cputime to minimize footprint.
---
 include/linux/cgroup-defs.h |  1 +
 kernel/cgroup/rstat.c       | 16 +++++++++++++---
 2 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/include/linux/cgroup-defs.h b/include/linux/cgroup-defs.h
index ae04035b6cbe..a2fcb3db6c52 100644
--- a/include/linux/cgroup-defs.h
+++ b/include/linux/cgroup-defs.h
@@ -315,6 +315,7 @@ struct cgroup_base_stat {
 #ifdef CONFIG_SCHED_CORE
 	u64 forceidle_sum;
 #endif
+	u64 ntime;
 };
 
 /*
diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
index a06b45272411..a77ba9a83bab 100644
--- a/kernel/cgroup/rstat.c
+++ b/kernel/cgroup/rstat.c
@@ -444,6 +444,7 @@ static void cgroup_base_stat_add(struct cgroup_base_stat *dst_bstat,
 #ifdef CONFIG_SCHED_CORE
 	dst_bstat->forceidle_sum += src_bstat->forceidle_sum;
 #endif
+	dst_bstat->ntime += src_bstat->ntime;
 }
 
 static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
@@ -455,6 +456,7 @@ static void cgroup_base_stat_sub(struct cgroup_base_stat *dst_bstat,
 #ifdef CONFIG_SCHED_CORE
 	dst_bstat->forceidle_sum -= src_bstat->forceidle_sum;
 #endif
+	dst_bstat->ntime -= src_bstat->ntime;
 }
 
 static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu)
@@ -535,7 +537,10 @@ void __cgroup_account_cputime_field(struct cgroup *cgrp,
 
 	switch (index) {
 	case CPUTIME_USER:
+		rstatc->bstat.cputime.utime += delta_exec;
+		break;
 	case CPUTIME_NICE:
+		rstatc->bstat.ntime += delta_exec;
 		rstatc->bstat.cputime.utime += delta_exec;
 		break;
 	case CPUTIME_SYSTEM:
@@ -591,6 +596,7 @@ static void root_cgroup_cputime(struct cgroup_base_stat *bstat)
 #ifdef CONFIG_SCHED_CORE
 		bstat->forceidle_sum += cpustat[CPUTIME_FORCEIDLE];
 #endif
+		bstat->ntime += cpustat[CPUTIME_NICE];
 	}
 }
 
@@ -608,13 +614,14 @@ static void cgroup_force_idle_show(struct seq_file *seq, struct cgroup_base_stat
 void cgroup_base_stat_cputime_show(struct seq_file *seq)
 {
 	struct cgroup *cgrp = seq_css(seq)->cgroup;
-	u64 usage, utime, stime;
+	u64 usage, utime, stime, ntime;
 
 	if (cgroup_parent(cgrp)) {
 		cgroup_rstat_flush_hold(cgrp);
 		usage = cgrp->bstat.cputime.sum_exec_runtime;
 		cputime_adjust(&cgrp->bstat.cputime, &cgrp->prev_cputime,
 			       &utime, &stime);
+		ntime = cgrp->bstat.ntime;
 		cgroup_rstat_flush_release(cgrp);
 	} else {
 		/* cgrp->bstat of root is not actually used, reuse it */
@@ -622,16 +629,19 @@ void cgroup_base_stat_cputime_show(struct seq_file *seq)
 		usage = cgrp->bstat.cputime.sum_exec_runtime;
 		utime = cgrp->bstat.cputime.utime;
 		stime = cgrp->bstat.cputime.stime;
+		ntime = cgrp->bstat.ntime;
 	}
 
 	do_div(usage, NSEC_PER_USEC);
 	do_div(utime, NSEC_PER_USEC);
 	do_div(stime, NSEC_PER_USEC);
+	do_div(ntime, NSEC_PER_USEC);
 
 	seq_printf(seq, "usage_usec %llu\n"
 		   "user_usec %llu\n"
-		   "system_usec %llu\n",
-		   usage, utime, stime);
+			 "system_usec %llu\n"
+			 "nice_usec %llu\n",
+			 usage, utime, stime, ntime);
 
 	cgroup_force_idle_show(seq, &cgrp->bstat);
 }
-- 
2.43.5

Re: [PATCH 1/2] Tracking cgroup-level niced CPU time

[Tejun Heo]

Hello,

On Fri, Aug 23, 2024 at 01:05:17PM -0700, JoshuaHahnjoshua.hahn6@gmail.com wrote:
> From: Joshua Hahn <joshua.hahn6@gmail.com>
> 
> Cgroup-level CPU statistics currently include time spent on
> user/system processes, but do not include niced CPU time (despite
> already being tracked). This patch exposes niced CPU time to the
> userspace, allowing users to get a better understanding of their
> hardware limits and can facilitate better load-balancing.

You aren't talking about the in-kernel scheduler's load balancer, right? If
so, can you please update the description? This is a bit too confusing for a
commit message for a kernel commit.

Thanks.

-- 
tejun

Re: [PATCH 1/2] Tracking cgroup-level niced CPU time

[Joshua Hahn]

Hello, thank you for reviewing the patch.

> > Cgroup-level CPU statistics currently include time spent on
> > user/system processes, but do not include niced CPU time (despite
> > already being tracked). This patch exposes niced CPU time to the
> > userspace, allowing users to get a better understanding of their
> > hardware limits and can facilitate better load-balancing.
>
> You aren't talking about the in-kernel scheduler's load balancer, right? If
> so, can you please update the description? This is a bit too confusing for a
> commit message for a kernel commit.

Thank you for pointing this out -- I'll edit the commit message to the
following in a v2:

Cgroup-level CPU statistics currently include time spent on
user/system processes, but do not include niced CPU time (despite
already being tracked). This patch exposes niced CPU time to the
userspace, allowing users to get a better understanding of their
hardware limits and can facilitate more informed workload distribution.

Thanks,
Joshua

[PATCH 2/2] Selftests for niced CPU statistics

[Joshua]

From: Joshua Hahn <joshua.hahn6@gmail.com>

Creates a cgroup with a single nice CPU hog process running.
fork() is called to generate the nice process because un-nicing is
not possible (see man nice(3)). If fork() was not used to generate
the CPU hog, we would run the rest of the cgroup selftest suite as a
nice process.
---
 tools/testing/selftests/cgroup/test_cpu.c | 72 +++++++++++++++++++++++
 1 file changed, 72 insertions(+)

diff --git a/tools/testing/selftests/cgroup/test_cpu.c b/tools/testing/selftests/cgroup/test_cpu.c
index dad2ed82f3ef..cd5550391f49 100644
--- a/tools/testing/selftests/cgroup/test_cpu.c
+++ b/tools/testing/selftests/cgroup/test_cpu.c
@@ -8,6 +8,7 @@
 #include <pthread.h>
 #include <stdio.h>
 #include <time.h>
+#include <unistd.h>
 
 #include "../kselftest.h"
 #include "cgroup_util.h"
@@ -229,6 +230,76 @@ static int test_cpucg_stats(const char *root)
 	return ret;
 }
 
+/*
+ * Creates a nice process that consumes CPU and checks that the elapsed
+ * usertime in the cgroup is close to the expected time.
+ */
+static int test_cpucg_nice(const char *root)
+{
+	int ret = KSFT_FAIL;
+	int status;
+	long user_usec, nice_usec;
+	long usage_seconds = 2;
+	long expected_nice_usec = usage_seconds * USEC_PER_SEC;
+	char *cpucg;
+	pid_t pid;
+
+	cpucg = cg_name(root, "cpucg_test");
+	if (!cpucg)
+		goto cleanup;
+
+	if (cg_create(cpucg))
+		goto cleanup;
+
+	user_usec = cg_read_key_long(cpucg, "cpu.stat", "user_usec");
+	nice_usec = cg_read_key_long(cpucg, "cpu.stat", "nice_usec");
+	if (user_usec != 0 || nice_usec != 0)
+		goto cleanup;
+
+	/*
+	 * We fork here to create a new process that can be niced without
+	 * polluting the nice value of other selftests
+	 */
+	pid = fork();
+	if (pid < 0) {
+		goto cleanup;
+	} else if (pid == 0) {
+		struct cpu_hog_func_param param = {
+			.nprocs = 1,
+			.ts = {
+				.tv_sec = usage_seconds,
+				.tv_nsec = 0,
+			},
+			.clock_type = CPU_HOG_CLOCK_PROCESS,
+		};
+
+		/* Try to keep niced CPU usage as constrained to hog_cpu as possible */
+		nice(1);
+		cg_run(cpucg, hog_cpus_timed, (void *)&param);
+		exit(0);
+	} else {
+		waitpid(pid, &status, 0);
+		if (!WIFEXITED(status))
+			goto cleanup;
+
+		user_usec = cg_read_key_long(cpucg, "cpu.stat", "user_usec");
+		nice_usec = cg_read_key_long(cpucg, "cpu.stat", "nice_usec");
+		if (nice_usec > user_usec || user_usec <= 0)
+			goto cleanup;
+
+		if (!values_close(nice_usec, expected_nice_usec, 1))
+			goto cleanup;
+
+		ret = KSFT_PASS;
+	}
+
+cleanup:
+	cg_destroy(cpucg);
+	free(cpucg);
+
+	return ret;
+}
+
 static int
 run_cpucg_weight_test(
 		const char *root,
@@ -686,6 +757,7 @@ struct cpucg_test {
 } tests[] = {
 	T(test_cpucg_subtree_control),
 	T(test_cpucg_stats),
+	T(test_cpucg_nice),
 	T(test_cpucg_weight_overprovisioned),
 	T(test_cpucg_weight_underprovisioned),
 	T(test_cpucg_nested_weight_overprovisioned),
-- 
2.43.5

Re: [PATCH 0/2] Exposing nice CPU usage to userspace

[Michal Koutný]

Hello.

On Fri, Aug 23, 2024 at 01:05:16PM GMT, JoshuaHahnjoshua.hahn6@gmail.com wrote:
> Niced CPU usage is a metric reported in host-level /proc/stat, but is
> not reported in cgroup-level statistics in cpu.stat. However, when a
> host contains multiple tasks across different workloads, it becomes
> difficult to gauage how much of the task is being spent on niced
> processes based on /proc/stat alone, since host-level metrics do not
> provide this cgroup-level granularity.

The difference between the two metrics is in cputime.c:
        index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;

> Exposing this metric will allow load balancers to correctly probe the
> niced CPU metric for each workload, and make more informed decisions
> when directing higher priority tasks.

How would this work? (E.g. if too little nice time -> reduce priority
of high prio tasks?)

Thanks,
Michal

Re: [PATCH 0/2] Exposing nice CPU usage to userspace

[Joshua Hahn]

Hello, thank you for reviewing the patch.

On Mon, Aug 26, 2024 at 10:43 AM Michal Koutný <mkoutny@suse.com> wrote:
> The difference between the two metrics is in cputime.c:
>         index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
>
> > Exposing this metric will allow load balancers to correctly probe the
> > niced CPU metric for each workload, and make more informed decisions
> > when directing higher priority tasks.
>
> How would this work? (E.g. if too little nice time -> reduce priority
> of high prio tasks?)

We can find what fraction of the task is being run as a nice process by
dividing the two metrics (nice / user) and determining the fraction of
niceness. When a high prio task comes into the load balancer and must
decide where the task should be delegated to, the balancer can use how much
of the task is nice as one factor in making the decision.

The reverse is also true; host-level information in /proc/stat may indicate that
a high percentage of CPU time is being used by nice processes, giving an
illusion that all tasks within the host are running nice processes,
when in reality,
it is just one task that is using a lot of nice CPU time, and other tasks are
running nonnice tasks. By including cgroup-level nice statistics, we can get
a clearer picture and avoid overloading a host with too many high prio tasks.

Like you suggested, this information can also help in re-prioritizing
the processes, which may help high prio tasks become executed quicker.

Thanks,
Joshua