37b995f6e7
It's been superseded by the atomic helpers.
The use of the atomic helpers provides a significant performance and scalability
improvement. Below is the result of running the atomic_add-test microbenchmark with:
$ x86_64-linux-user/qemu-x86_64 tests/atomic_add-bench -o 5000000 -r $r -n $n
, where $n is the number of threads and $r is the allowed range for the additions.
The scenarios measured are:
- atomic: implements x86' ADDL with the atomic_add helper (i.e. this patchset)
- cmpxchg: implement x86' ADDL with a TCG loop using the cmpxchg helper
- master: before this patchset
Results sorted in ascending range, i.e. descending degree of contention.
Y axis is Throughput in Mops/s. Tests are run on an AMD machine with 64
Opteron 6376 cores.
atomic_add-bench: 5000000 ops/thread, [0,1] range
25 ++---------+----------+---------+----------+----------+----------+---++
+ atomic +-E--+ + + + + + |
|cmpxchg +-H--+ |
20 +Emaster +-N--+ ++
|| |
|++ |
|| |
15 +++ ++
|N| |
|+| |
10 ++| ++
|+|+ |
| | -+E+------ +++ ---+E+------+E+------+E+-----+E+------+E|
|+E+E+- +++ +E+------+E+-- |
5 ++|+ ++
|+N+H+--- +++ |
++++N+--+H++----+++ + +++ --++H+------+H+------+H++----+H+---+--- |
0 ++---------+-----H----+---H-----+----------+----------+----------+---H+
0 10 20 30 40 50 60
Number of threads
atomic_add-bench: 5000000 ops/thread, [0,2] range
25 ++---------+----------+---------+----------+----------+----------+---++
++atomic +-E--+ + + + + + |
|cmpxchg +-H--+ |
20 ++master +-N--+ ++
|E| |
|++ |
||E |
15 ++| ++
|N|| |
|+|| ---+E+------+E+-----+E+------+E|
10 ++| | ---+E+------+E+-----+E+--- +++ +++
||H+E+--+E+-- |
|+++++ |
| || |
5 ++|+H+-- +++ ++
|+N+ - ---+H+------+H+------ |
+ +N+--+H++----+H+---+--+H+----++H+--- + + +H+---+--+H|
0 ++---------+----------+---------+----------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench: 5000000 ops/thread, [0,8] range
40 ++---------+----------+---------+----------+----------+----------+---++
++atomic +-E--+ + + + + + |
35 +cmpxchg +-H--+ ++
| master +-N--+ ---+E+------+E+------+E+-----+E+------+E|
30 ++| ---+E+-- +++ ++
| | -+E+--- |
25 ++E ---- +++ ++
|+++++ -+E+ |
20 +E+ E-- +++ ++
|H|+++ |
|+| +H+------- |
15 ++H+ ---+++ +H+------ ++
|N++H+-- +++--- +H+------++|
10 ++ +++ - +++ ---+H+ +++ +H+
| | +H+-----+H+------+H+-- |
5 ++| +++ ++
++N+N+--+N++ + + + + + |
0 ++---------+----------+---------+----------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench: 5000000 ops/thread, [0,128] range
160 ++---------+---------+----------+---------+----------+----------+---++
+ atomic +-E--+ + + + + + |
140 +cmpxchg +-H--+ +++ +++ ++
| master +-N--+ E--------E------+E+------++|
120 ++ --| | +++ E+
| -- +++ +++ ++|
100 ++ - ++
| +++- +++ ++|
80 ++ -+E+ -+H+------+H+------H--------++
| ---- ---- +++ H|
| ---+E+-----+E+- ---+H+ ++|
60 ++ +E+--- +++ ---+H+--- ++
| --+++ ---+H+-- |
40 ++ +E+-+H+--- ++
| +H+ |
20 +EE+ ++
+N+ + + + + + + |
0 ++N-N---N--+---------+----------+---------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
atomic_add-bench: 5000000 ops/thread, [0,1024] range
350 ++---------+---------+----------+---------+----------+----------+---++
+ atomic +-E--+ + + + + + |
300 +cmpxchg +-H--+ +++
| master +-N--+ +++ ||
| +++ | ----E|
250 ++ | ----E---- ++
| ----E--- | ---+H|
200 ++ -+E+--- +++ ---+H+--- ++
| ---- -+H+-- |
| +E+ +++ ---- +++ |
150 ++ ---+++ ---+H+- ++
| --- -+H+-- |
100 ++ ---+E+ ---- +++ ++
| +++ ---+E+-----+H+- |
| -+E+------+H+-- |
50 ++ +E+ ++
+EE+ + + + + + + |
0 ++N-N---N--+---------+----------+---------+----------+----------+---++
0 10 20 30 40 50 60
Number of threads
hi-res: http://imgur.com/a/fMRmq
For master I stopped measuring master after 8 threads, because there is little
point in measuring the well-known performance collapse of a contended lock.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Message-Id: <1467054136-10430-21-git-send-email-cota@braap.org>
Signed-off-by: Richard Henderson <rth@twiddle.net>
|
||
|---|---|---|
| .. | ||
| arch_dump.c | ||
| arch_memory_mapping.c | ||
| bpt_helper.c | ||
| cc_helper_template.h | ||
| cc_helper.c | ||
| cpu-qom.h | ||
| cpu.c | ||
| cpu.h | ||
| excp_helper.c | ||
| fpu_helper.c | ||
| gdbstub.c | ||
| helper.c | ||
| helper.h | ||
| hyperv.c | ||
| hyperv.h | ||
| int_helper.c | ||
| kvm_i386.h | ||
| kvm-stub.c | ||
| kvm.c | ||
| machine.c | ||
| Makefile.objs | ||
| mem_helper.c | ||
| misc_helper.c | ||
| monitor.c | ||
| mpx_helper.c | ||
| ops_sse_header.h | ||
| ops_sse.h | ||
| seg_helper.c | ||
| shift_helper_template.h | ||
| smm_helper.c | ||
| svm_helper.c | ||
| svm.h | ||
| TODO | ||
| trace-events | ||
| translate.c | ||