tools/cpuunclaimed.py

#!/usr/bin/env python
# @lint-avoid-python-3-compatibility-imports
#
# cpuunclaimed   Sample CPU run queues and calculate unclaimed idle CPU.
#                For Linux, uses BCC, eBPF.
#
# This samples the length of the run queues and determine when there are idle
# CPUs, yet queued threads waiting their turn. Report the amount of idle
# (yet unclaimed by waiting threads) CPU as a system-wide percentage.
#
# This situation can happen for a number of reasons:
#
# - An application has been bound to some, but not all, CPUs, and has runnable
#   threads that cannot migrate to other CPUs due to this configuration.
# - CPU affinity: an optimization that leaves threads on CPUs where the CPU
#   caches are warm, even if this means short periods of waiting while other
#   CPUs are idle. The wait period is tunale (see sysctl, kernel.sched*).
# - Scheduler bugs.
#
# An unclaimed idle of < 1% is likely to be CPU affinity, and not usually a
# cause for concern. By leaving the CPU idle, overall throughput of the system
# may be improved. This tool is best for identifying larger issues, > 2%, due
# to the coarseness of its 99 Hertz samples.
#
# This is an experimental tool that currently works by use of sampling to
# keep overheads low. Tool assumptions:
#
# - CPU samples consistently fire around the same offset. There will sometimes
#   be a lag as a sample is delayed by higher-priority interrupts, but it is
#   assumed the subsequent samples will catch up to the expected offsets (as
#   is seen in practice). You can use -J to inspect sample offsets. Some
#   systems can power down CPUs when idle, and when they wake up again they
#   may begin firing at a skewed offset: this tool will detect the skew, print
#   an error, and exit.
# - All CPUs are online (see ncpu).
#
# If this identifies unclaimed CPU, you can double check it by dumping raw
# samples (-j), as well as using other tracing tools to instrument scheduler
# events (although this latter approach has much higher overhead).
#
# This tool passes all sampled events to user space for post processing.
# I originally wrote this to do the calculations entirerly in kernel context,
# and only pass a summary. That involves a number of challenges, and the
# overhead savings may not outweigh the caveats. You can see my WIP here:
# https://gist.github.com/brendangregg/731cf2ce54bf1f9a19d4ccd397625ad9
#
# USAGE: cpuunclaimed [-h] [-j] [-J] [-T] [interval] [count]
#
# If you see "Lost 1881 samples" warnings, try increasing wakeup_hz.
#
# REQUIRES: Linux 4.9+ (BPF_PROG_TYPE_PERF_EVENT support). Under tools/old is
# a version of this tool that may work on Linux 4.6 - 4.8.
#
# Copyright 2016 Netflix, Inc.
# Licensed under the Apache License, Version 2.0 (the "License")
#
# 20-Dec-2016   Brendan Gregg   Created this.

from __future__ import print_function
from bcc import BPF, PerfType, PerfSWConfig
from time import sleep, strftime
from ctypes import c_int
import argparse
import multiprocessing
from os import getpid, system
import ctypes as ct

# arguments
examples = """examples:
    ./cpuunclaimed            # sample and calculate unclaimed idle CPUs,
                              # output every 1 second (default)
    ./cpuunclaimed 5 10       # print 5 second summaries, 10 times
    ./cpuunclaimed -T 1       # 1s summaries and timestamps
    ./cpuunclaimed -j         # raw dump of all samples (verbose), CSV
"""
parser = argparse.ArgumentParser(
    description="Sample CPU run queues and calculate unclaimed idle CPU",
    formatter_class=argparse.RawDescriptionHelpFormatter,
    epilog=examples)
parser.add_argument("-j", "--csv", action="store_true",
    help="print sample summaries (verbose) as comma-separated values")
parser.add_argument("-J", "--fullcsv", action="store_true",
    help="print sample summaries with extra fields: CPU sample offsets")
parser.add_argument("-T", "--timestamp", action="store_true",
    help="include timestamp on output")
parser.add_argument("interval", nargs="?", default=-1,
    help="output interval, in seconds")
parser.add_argument("count", nargs="?", default=99999999,
    help="number of outputs")
parser.add_argument("--ebpf", action="store_true",
    help=argparse.SUPPRESS)
args = parser.parse_args()
countdown = int(args.count)
frequency = 99
dobind = 1
wakeup_hz = 10                      # frequency to read buffers
wakeup_s = float(1) / wakeup_hz
ncpu = multiprocessing.cpu_count()  # assume all are online
debug = 0

# process arguments
if args.fullcsv:
    args.csv = True
if args.csv:
    interval = 0.2
if args.interval != -1 and (args.fullcsv or args.csv):
    print("ERROR: cannot use interval with either -j or -J. Exiting.")
    exit()
if args.interval == -1:
    args.interval = "1"
interval = float(args.interval)

# define BPF program
bpf_text = """
#include <uapi/linux/ptrace.h>
#include <uapi/linux/bpf_perf_event.h>
#include <linux/sched.h>

struct data_t {
    u64 ts;
    u64 cpu;
    u64 len;
};

BPF_PERF_OUTPUT(events);

// Declare enough of cfs_rq to find nr_running, since we can't #import the
// header. This will need maintenance. It is from kernel/sched/sched.h:
struct cfs_rq_partial {
    struct load_weight load;
    unsigned int nr_running, h_nr_running;
};

int do_perf_event(struct bpf_perf_event_data *ctx)
{
    int cpu = bpf_get_smp_processor_id();
    u64 now = bpf_ktime_get_ns();

    /*
     * Fetch the run queue length from task->se.cfs_rq->nr_running. This is an
     * unstable interface and may need maintenance. Perhaps a future version
     * of BPF will support task_rq(p) or something similar as a more reliable
     * interface.
     */
    unsigned int len = 0;
    struct task_struct *task = NULL;
    struct cfs_rq_partial *my_q = NULL;
    task = (struct task_struct *)bpf_get_current_task();
    my_q = (struct cfs_rq_partial *)task->se.cfs_rq;
    len = my_q->nr_running;

    struct data_t data = {.ts = now, .cpu = cpu, .len = len};
    events.perf_submit(ctx, &data, sizeof(data));

    return 0;
}
"""

# code substitutions
if debug or args.ebpf:
    print(bpf_text)
    if args.ebpf:
        exit()

# initialize BPF & perf_events
b = BPF(text=bpf_text)
# TODO: check for HW counters first and use if more accurate
b.attach_perf_event(ev_type=PerfType.SOFTWARE,
    ev_config=PerfSWConfig.TASK_CLOCK, fn_name="do_perf_event",
    sample_period=0, sample_freq=frequency)

if args.csv:
    if args.timestamp:
        print("TIME", end=",")
    print("TIMESTAMP_ns", end=",")
    print(",".join("CPU" + str(c) for c in range(ncpu)), end="")
    if args.fullcsv:
        print(",", end="")
        print(",".join("OFFSET_ns_CPU" + str(c) for c in range(ncpu)), end="")
    print()
else:
    print(("Sampling run queues... Output every %s seconds. " +
          "Hit Ctrl-C to end.") % args.interval)
class Data(ct.Structure):
    _fields_ = [
        ("ts", ct.c_ulonglong),
        ("cpu", ct.c_ulonglong),
        ("len", ct.c_ulonglong)
    ]

samples = {}
group = {}
last = 0

# process event
def print_event(cpu, data, size):
    event = ct.cast(data, ct.POINTER(Data)).contents
    samples[event.ts] = {}
    samples[event.ts]['cpu'] = event.cpu
    samples[event.ts]['len'] = event.len

exiting = 0 if args.interval else 1
slept = float(0)

# Choose the elapsed time from one sample group to the next that identifies a
# new sample group (a group being a set of samples from all CPUs). The
# earliest timestamp is compared in each group. This trigger is also used
# for sanity testing, if a group's samples exceed half this value.
trigger = int(0.8 * (1000000000 / frequency))

# read events
b["events"].open_perf_buffer(print_event, page_cnt=64)
while 1:
    # allow some buffering by calling sleep(), to reduce the context switch
    # rate and lower overhead.
    try:
        if not exiting:
            sleep(wakeup_s)
    except KeyboardInterrupt:
        exiting = 1
    b.perf_buffer_poll()
    slept += wakeup_s

    if slept < 0.999 * interval:   # floating point workaround
        continue
    slept = 0

    positive = 0  # number of samples where an idle CPU could have run work
    running = 0
    idle = 0
    if debug >= 2:
        print("DEBUG: begin samples loop, count %d" % len(samples))
    for e in sorted(samples):
        if debug >= 2:
            print("DEBUG: ts %d cpu %d len %d delta %d trig %d" % (e,
                  samples[e]['cpu'], samples[e]['len'], e - last,
                  e - last > trigger))

        # look for time jumps to identify a new sample group
        if e - last > trigger:

            # first first group timestamp, and sanity test
            g_time = 0
            g_max = 0
            for ge in sorted(group):
                if g_time == 0:
                    g_time = ge
                g_max = ge

            # process previous sample group
            if args.csv:
                lens = [0] * ncpu
                offs = [0] * ncpu
                for ge in sorted(group):
                    lens[samples[ge]['cpu']] = samples[ge]['len']
                    if args.fullcsv:
                        offs[samples[ge]['cpu']] = ge - g_time
                if g_time > 0:      # else first sample
                    if args.timestamp:
                        print("%-8s" % strftime("%H:%M:%S"), end=",")
                    print("%d" % g_time, end=",")
                    print(",".join(str(lens[c]) for c in range(ncpu)), end="")
                    if args.fullcsv:
                        print(",", end="")
                        print(",".join(str(offs[c]) for c in range(ncpu)))
                    else:
                        print()
            else:
                # calculate stats
                g_running = 0
                g_queued = 0
                for ge in group:
                    if samples[ge]['len'] > 0:
                        g_running += 1
                    if samples[ge]['len'] > 1:
                        g_queued += samples[ge]['len'] - 1
                g_idle = ncpu - g_running

                # calculate the number of threads that could have run as the
                # minimum of idle and queued
                if g_idle > 0 and g_queued > 0:
                    if g_queued > g_idle:
                        i = g_idle
                    else:
                        i = g_queued
                    positive += i
                running += g_running
                idle += g_idle

            # now sanity test, after -J output
            g_range = g_max - g_time
            if g_range > trigger / 2:
                # if a sample group exceeds half the interval, we can no
                # longer draw conclusions about some CPUs idle while others
                # have queued work. Error and exit. This can happen when
                # CPUs power down, then start again on different offsets.
                # TODO: Since this is a sampling tool, an error margin should
                # be anticipated, so an improvement may be to bump a counter
                # instead of exiting, and only exit if this counter shows
                # a skewed sample rate of over, say, 1%. Such an approach
                # would allow a small rate of outliers (sampling error),
                # and, we could tighten the trigger to be, say, trigger / 5.
                # In the case of a power down, if it's detectable, perhaps
                # the tool could reinitialize the timers (although exiting
                # is simple and works).
                print(("ERROR: CPU samples arrived at skewed offsets " +
                      "(CPUs may have powered down when idle), " +
                      "spanning %d ns (expected < %d ns). Debug with -J, " +
                      "and see the man page. As output may begin to be " +
                      "unreliable, exiting.") % (g_range, trigger / 2))
                exit()

            # these are done, remove
            for ge in sorted(group):
                del samples[ge]

            # begin next group
            group = {}
            last = e

        # stash this timestamp in a sample group dict
        group[e] = 1

    if not args.csv:
        total = running + idle
        unclaimed = util = 0

        if debug:
            print("DEBUG: hit %d running %d idle %d total %d buffered %d" % (
                  positive, running, idle, total, len(samples)))

        if args.timestamp:
            print("%-8s " % strftime("%H:%M:%S"), end="")

        # output
        if total:
            unclaimed = float(positive) / total
            util = float(running) / total
        print("%%CPU %6.2f%%, unclaimed idle %0.2f%%" % (100 * util,
              100 * unclaimed))

    countdown -= 1
    if exiting or countdown == 0:
        exit()