synapse.metrics: implement detailed memory usage reporting on PyPy (#7536)

PyPy's gc.get_stats() returns an object containing detailed allocator statistics
which could be beneficial to collect as metrics.

Signed-off-by: Ivan Shapovalov <intelfx@intelfx.name>
This commit is contained in:
Ivan Shapovalov 2020-05-22 13:08:41 +03:00 committed by GitHub
parent 8c75da916c
commit ac481a738e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 79 additions and 1 deletions

1
changelog.d/7536.misc Normal file
View File

@ -0,0 +1 @@
Synapse now exports [detailed allocator statistics](https://doc.pypy.org/en/latest/gc_info.html#gc-get-stats) and basic GC timings as Prometheus metrics (`pypy_gc_time_seconds_total` and `pypy_memory_bytes`) when run under PyPy. Contributed by Ivan Shapovalov.

View File

@ -26,7 +26,12 @@ import six
import attr
from prometheus_client import Counter, Gauge, Histogram
from prometheus_client.core import REGISTRY, GaugeMetricFamily, HistogramMetricFamily
from prometheus_client.core import (
REGISTRY,
CounterMetricFamily,
GaugeMetricFamily,
HistogramMetricFamily,
)
from twisted.internet import reactor
@ -338,6 +343,78 @@ class GCCounts(object):
if not running_on_pypy:
REGISTRY.register(GCCounts())
#
# PyPy GC / memory metrics
#
class PyPyGCStats(object):
def collect(self):
# @stats is a pretty-printer object with __str__() returning a nice table,
# plus some fields that contain data from that table.
# unfortunately, fields are pretty-printed themselves (i. e. '4.5MB').
stats = gc.get_stats(memory_pressure=False) # type: ignore
# @s contains same fields as @stats, but as actual integers.
s = stats._s # type: ignore
# also note that field naming is completely braindead
# and only vaguely correlates with the pretty-printed table.
# >>>> gc.get_stats(False)
# Total memory consumed:
# GC used: 8.7MB (peak: 39.0MB) # s.total_gc_memory, s.peak_memory
# in arenas: 3.0MB # s.total_arena_memory
# rawmalloced: 1.7MB # s.total_rawmalloced_memory
# nursery: 4.0MB # s.nursery_size
# raw assembler used: 31.0kB # s.jit_backend_used
# -----------------------------
# Total: 8.8MB # stats.memory_used_sum
#
# Total memory allocated:
# GC allocated: 38.7MB (peak: 41.1MB) # s.total_allocated_memory, s.peak_allocated_memory
# in arenas: 30.9MB # s.peak_arena_memory
# rawmalloced: 4.1MB # s.peak_rawmalloced_memory
# nursery: 4.0MB # s.nursery_size
# raw assembler allocated: 1.0MB # s.jit_backend_allocated
# -----------------------------
# Total: 39.7MB # stats.memory_allocated_sum
#
# Total time spent in GC: 0.073 # s.total_gc_time
pypy_gc_time = CounterMetricFamily(
"pypy_gc_time_seconds_total", "Total time spent in PyPy GC", labels=[],
)
pypy_gc_time.add_metric([], s.total_gc_time / 1000)
yield pypy_gc_time
pypy_mem = GaugeMetricFamily(
"pypy_memory_bytes",
"Memory tracked by PyPy allocator",
labels=["state", "class", "kind"],
)
# memory used by JIT assembler
pypy_mem.add_metric(["used", "", "jit"], s.jit_backend_used)
pypy_mem.add_metric(["allocated", "", "jit"], s.jit_backend_allocated)
# memory used by GCed objects
pypy_mem.add_metric(["used", "", "arenas"], s.total_arena_memory)
pypy_mem.add_metric(["allocated", "", "arenas"], s.peak_arena_memory)
pypy_mem.add_metric(["used", "", "rawmalloced"], s.total_rawmalloced_memory)
pypy_mem.add_metric(["allocated", "", "rawmalloced"], s.peak_rawmalloced_memory)
pypy_mem.add_metric(["used", "", "nursery"], s.nursery_size)
pypy_mem.add_metric(["allocated", "", "nursery"], s.nursery_size)
# totals
pypy_mem.add_metric(["used", "totals", "gc"], s.total_gc_memory)
pypy_mem.add_metric(["allocated", "totals", "gc"], s.total_allocated_memory)
pypy_mem.add_metric(["used", "totals", "gc_peak"], s.peak_memory)
pypy_mem.add_metric(["allocated", "totals", "gc_peak"], s.peak_allocated_memory)
yield pypy_mem
if running_on_pypy:
REGISTRY.register(PyPyGCStats())
#
# Twisted reactor metrics
#