[PATCH v3 0/3] perf record: adapt NUMA awareness to machines with #CPUs > 1K

From: Alexey Budankov
Date: Fri Nov 29 2019 - 05:01:09 EST



Current implementation of cpu_set_t type by glibc has internal cpu
mask size limitation of no more than 1024 CPUs. This limitation confines
NUMA awareness of Perf tool in record mode, thru --affinity option,
to the first 1024 CPUs on machines with larger amount of CPUs.

This patch set enables Perf tool to overcome 1024 CPUs limitation by
using a dedicated struct mmap_cpu_mask type and applying tool's bitmap
API operations to manipulate affinity masks of the tool's thread and
the mmaped data buffers.

tools bitmap API has been extended with bitmap_free() function and
bitmap_equal() operation whose implementation is derived from the
kernel one.

---
Alexey Budankov (3):
tools bitmap: implement bitmap_equal() operation at bitmap API
perf mmap: declare type for cpu mask of arbitrary length
perf record: adapt affinity to machines with #CPUs > 1K

tools/include/linux/bitmap.h | 30 +++++++++++++++++++++++++++
tools/lib/bitmap.c | 15 ++++++++++++++
tools/perf/builtin-record.c | 27 ++++++++++++++++++------
tools/perf/util/mmap.c | 40 ++++++++++++++++++++++++++++++------
tools/perf/util/mmap.h | 13 +++++++++++-
5 files changed, 112 insertions(+), 13 deletions(-)

---
Changes in v3:
- implemented perf_mmap__print_cpu_mask() function
- use perf_mmap__print_cpu_mask() to log thread and mmap cpus masks
when verbose level is equal to 2
Changes in v2:
- implemented bitmap_free() for symmetry with bitmap_alloc()
- capitalized MMAP_CPU_MASK_BYTES() macro
- returned -1 from perf_mmap__setup_affinity_mask()
- implemented releasing of masks using bitmap_free()
- moved debug printing under -vv option

---
Testing:

tools/perf/perf record -vv --affinity=cpu -- ls
thread mask[8]: empty
...
------------------------------------------------------------
perf_event_attr:
...
mmap size 528384B
0x7fddecf760b8: mmap mask[8]: 0
0x7fddecf86180: mmap mask[8]: 1
0x7fddecf96248: mmap mask[8]: 2
0x7fddecfa6310: mmap mask[8]: 3
0x7fddecfb63d8: mmap mask[8]: 4
0x7fddecfc64a0: mmap mask[8]: 5
0x7fddecfd6568: mmap mask[8]: 6
0x7fddecfe6630: mmap mask[8]: 7
------------------------------------------------------------
perf_event_attr:
...
Synthesizing TSC conversion information
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
arch copy Documentation init kernel MAINTAINERS modules.builtin.modinfo perf.data scripts System.map vmlinux
block COPYING drivers ipc lbuild Makefile modules.order perf.data.old security tools vmlinux.o
certs CREDITS fs Kbuild lib mm Module.symvers README sound usr
config-5.2.7-100.fc29.x86_64 crypto include Kconfig LICENSES modules.builtin net samples stdio virt
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
[ perf record: Woken up 0 times to write data ]
0x7fddecf760b8: thread mask[8]: 0
0x7fddecf86180: thread mask[8]: 1
0x7fddecf96248: thread mask[8]: 2
0x7fddecfa6310: thread mask[8]: 3
0x7fddecfb63d8: thread mask[8]: 4
0x7fddecfc64a0: thread mask[8]: 5
0x7fddecfd6568: thread mask[8]: 6
0x7fddecfe6630: thread mask[8]: 7
Looking at the vmlinux_path (8 entries long)
Using vmlinux for symbols
[ perf record: Captured and wrote 0.014 MB perf.data (8 samples) ]

--
2.20.1