[lkml]   [2008]   [Nov]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [Bug #11308] tbench regression on each kernel release from 2.6.22 -> 2.6.28

* Linus Torvalds <> wrote:

> On Mon, 17 Nov 2008, Eric Dumazet wrote:
> > Ingo Molnar a écrit :
> > > it gives a small speedup of ~1% on my box:
> > >
> > > before: Throughput 3437.65 MB/sec 64 procs
> > > after: Throughput 3473.99 MB/sec 64 procs
> >
> > Strange, I get 2350 MB/sec on my 8 cpus box. "tbench 8"
> I think Ingo may have a Nehalem. Let's just say that those things
> rock, and have rather good memory throughput.

hm, i'm not sure whether i can post benchmarks from the Nehalem box -
but i can confirm it in general terms that it's rather nice ;-)

This was run on another testbox (4x4 Barcelona) that rocks similarly
well in terms of memory subsystem latencies: which seems to be
tbench's main current critical path.

For the tbench bragging rights i'd probably turn off CONFIG_SECURITY
and a few other options. Plus i'd run with 16 threads only - in this
test i ran with 4x overload (64 tbench threads, not 16) to stress the
scheduler harder.

Although we degrade very gently with overload so the numbers arent all
that much different:

16 threads: Throughput 3463.14 MB/sec 16 procs
64 threads: Throughput 3473.99 MB/sec 64 procs
256 threads: Throughput 3457.67 MB/sec 256 procs
1024 threads: Throughput 3448.85 MB/sec 1024 procs

[ so it's the same within noise range. ]

1024 threads is already a massive 64x overload so beyond any
reasonable limit of workload sanity.

Which suggests that the main limitation factor is cacheline ping-pong
that is already in full effect at 16 threads.

Which is supported by the "most expensive instructions" top-10 sorted

RIP #hits

[ usercopy ]
ffffffff80350fcd: 1373300 f3 48 a5 rep movsq %ds:(%rsi),%es:(%rdi)

ffffffff804a2f33: <sock_rfree>:
ffffffff804a2f34: 985253 48 89 e5 mov %rsp,%rbp

ffffffff804d2eb7: <ip_local_deliver>:
ffffffff804d2eb8: 432659 48 89 e5 mov %rsp,%rbp

ffffffff804aa23c: <constant_test_bit>: [ => napi_disable_pending() ]
ffffffff804aa24c: 374052 89 d1 mov %edx,%ecx

ffffffff804d5076: <ip_dont_fragment>:
ffffffff804d5076: 310051 8a 97 56 02 00 00 mov 0x256(%rdi),%dl

ffffffff804d9b17: <__inet_lookup_established>:
ffffffff804d9bdf: 247224 eb ba jmp ffffffff804d9b9b <__inet_lookup_established+0x84>

ffffffff80321529: <selinux_ip_postroute>:
ffffffff8032152a: 183700 48 89 e5 mov %rsp,%rbp

ffffffff8020c020: <system_call>:
ffffffff8020c020: 183600 0f 01 f8 swapgs

ffffffff8051884a: <netlbl_enabled>:
ffffffff8051884a: 179538 55 push %rbp

The usual profiling caveat applies: it's not _these_ instructions that
matter, but the surrounding code that calls them. Profiling overhead
is delayed by a couple of instructions - the more out-of-order a CPU
is, the larger this delay can be. But even a quick look to the list
above shows that all of the heavy cachemisses are generated by

Beyond the usual suspects of syscall entry and memcpy, it's only
networking. We dont even have the mov %cr3 TLB flush overhead in this
list, load_cr3() is a distant #30:

ffffffff8023049f: 0 0f 22 d8 mov %rax,%cr3
ffffffff802304a2: 126303 c9 leaveq

The place for the sock_rfree() hit looks a bit weird, and i'll
investigate it now a bit more to place the real overhead point
properly. (i already mapped the test-bit overhead: that comes from

The first entry is 10x the cost of the last entry in the list so
clearly we've got 1-2 brutal cacheline ping-pongs that dominate the
overhead of this workload.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2008-11-17 19:27    [W:0.146 / U:0.024 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site