@@ -14,13 +14,27 @@ class SystemStats:
bn = d.getVar('BUILDNAME')
bsdir = os.path.join(d.getVar('BUILDSTATS_BASE'), bn)
bb.utils.mkdirhier(bsdir)
+ file_handlers = [('diskstats', self._reduce_diskstats),
+ ('meminfo', self._reduce_meminfo),
+ ('stat', self._reduce_stat)]
+
+ # Some hosts like openSUSE have readable /proc/pressure files
+ # but throw errors when these files are opened. Catch these error
+ # and ensure that the reduce_proc_pressure directory is not created.
+ if os.path.exists("/proc/pressure"):
+ try:
+ source = open('/proc/pressure/cpu', 'rb')
+ source.read()
+ pressuredir = os.path.join(bsdir, 'reduced_proc_pressure')
+ bb.utils.mkdirhier(pressuredir)
+ file_handlers.extend([('pressure/cpu', self._reduce_pressure),
+ ('pressure/io', self._reduce_pressure),
+ ('pressure/memory', self._reduce_pressure)])
+ except Exception:
+ pass
self.proc_files = []
- for filename, handler in (
- ('diskstats', self._reduce_diskstats),
- ('meminfo', self._reduce_meminfo),
- ('stat', self._reduce_stat),
- ):
+ for filename, handler in (file_handlers):
# The corresponding /proc files might not exist on the host.
# For example, /proc/diskstats is not available in virtualized
# environments like Linux-VServer. Silently skip collecting
@@ -48,13 +62,15 @@ class SystemStats:
self.diskstats_ltime = None
self.diskstats_data = None
self.stat_ltimes = None
+ # Last time we sampled /proc/pressure. All resources stored in a single dict with the key as filename
+ self.last_pressure = {"pressure/cpu": None, "pressure/io": None, "pressure/memory": None}
def close(self):
self.monitor_disk.close()
for _, output, _ in self.proc_files:
output.close()
- def _reduce_meminfo(self, time, data):
+ def _reduce_meminfo(self, time, data, filename):
"""
Extracts 'MemTotal', 'MemFree', 'Buffers', 'Cached', 'SwapTotal', 'SwapFree'
and writes their values into a single line, in that order.
@@ -75,7 +91,7 @@ class SystemStats:
disk = linetokens[2]
return self.diskstats_regex.match(disk)
- def _reduce_diskstats(self, time, data):
+ def _reduce_diskstats(self, time, data, filename):
relevant_tokens = filter(self._diskstats_is_relevant_line, map(lambda x: x.split(), data.split(b'\n')))
diskdata = [0] * 3
reduced = None
@@ -104,10 +120,10 @@ class SystemStats:
return reduced
- def _reduce_nop(self, time, data):
+ def _reduce_nop(self, time, data, filename):
return (time, data)
- def _reduce_stat(self, time, data):
+ def _reduce_stat(self, time, data, filename):
if not data:
return None
# CPU times {user, nice, system, idle, io_wait, irq, softirq} from first line
@@ -126,6 +142,27 @@ class SystemStats:
self.stat_ltimes = times
return reduced
+ def _reduce_pressure(self, time, data, filename):
+ """
+ Return reduced pressure: {avg10, avg60, avg300} and delta total compared to the previous sample
+ for the cpu, io and memory resources. A common function is used for all 3 resources since the
+ format of the /proc/pressure file is the same in each case.
+ """
+ if not data:
+ return None
+ tokens = data.split(b'\n', 1)[0].split()
+ avg10 = float(tokens[1].split(b'=')[1])
+ avg60 = float(tokens[2].split(b'=')[1])
+ avg300 = float(tokens[3].split(b'=')[1])
+ total = int(tokens[4].split(b'=')[1])
+
+ reduced = None
+ if self.last_pressure[filename]:
+ delta = total - self.last_pressure[filename]
+ reduced = (time, (avg10, avg60, avg300, delta))
+ self.last_pressure[filename] = total
+ return reduced
+
def sample(self, event, force):
now = time.time()
if (now - self.last_proc > self.min_seconds) or force:
@@ -133,7 +170,7 @@ class SystemStats:
with open(os.path.join('/proc', filename), 'rb') as input:
data = input.read()
if handler:
- reduced = handler(now, data)
+ reduced = handler(now, data, filename)
else:
reduced = (now, data)
if reduced:
The Linux pressure monitoring system helps determine when system resources are being overutilized by measuring how contended the CPU, IO and memory are. This information can be found under /proc/pressure/ which contains 3 files - cpu, memory and io. In each of the files, the format is as follows: some avg10=70.24 avg60=68.52 avg300=69.91 total=3559632828 full avg10=57.59 avg60=58.06 avg300=60.38 total=3300487258 The "some" state of a given resource represents when one or more tasks are delayed on that resource whereas the "full" state represents when all the tasks are delayed. Currently, we only collect data from the "some" state but the "full" data can simply be appended to the log files if neccessary. The "avg10", "avg60" and "avg300" fields represent the average percentage of time runnable tasks were delayed in the last 10, 60 or 300 seconds respectively. The "total" field represents the total time, in microseconds, that some runnable task was delayed on a resource. More information can be found at: https://www.kernel.org/doc/html/latest/accounting/psi.html and in the source code under kernel/sched/psi.c This commit adds functionality to collect and log the "some" CPU, memory and IO pressure. The "avg10", "avg60" and "avg300" fields are logged without change. In place of the "total" field, the difference between the current "total" and the previous sample's "total" is logged, allowing the measurement of pressure in between each polling interval, as was done for /proc/stat data. The log files are stored in: <build_name>/tmp/buildstats/<build_time>/reduced_proc_pressure/{cpu,io,memory}.log mirroring the directory structure of /proc/pressure. If the /proc/pressure directory does not exist or the resource files can't be read/opened, the reduced_proc_pressure directory is not created. Signed-off-by: Aryaman Gupta <aryaman.gupta@windriver.com> Signed-off-by: Randy MacLeod <randy.macleod@windriver.com> --- meta/lib/buildstats.py | 57 ++++++++++++++++++++++++++++++++++-------- 1 file changed, 47 insertions(+), 10 deletions(-)