MIME-Version: 1.0
In-Reply-To: <20160308173709.GC3017@gmail.com>
References: <CACT4Y+Y6+S9E+vJUZzWxyU+0M5i3L_+8knhxFtfaWYkdN5Vusg@mail.gmail.com>
 <20160308155423.GA16587@gmail.com> <CACT4Y+YwsGU1t62o20GTNQZ1=uQW1CateC+gSRFMG7Pmxi+WQw@mail.gmail.com>
 <20160308162703.GB30211@gmail.com> <CACT4Y+aTmdQfeBcQ=8dt=XV7MBHP-+kDLbXAZ_cJ93npGrih1g@mail.gmail.com>
 <20160308164438.GA24109@gmail.com> <20160308164859.GA27516@gmail.com>
 <CACT4Y+amJpRg9jdz1vEupJO8gsbeh7PuOKTiuvkuu7cWigY50Q@mail.gmail.com>
 <20160308172425.GA3017@gmail.com> <CACT4Y+ZbJk6smr11ctniFW577_BKcjHJd1duyA4h7taPmAPP=A@mail.gmail.com>
 <20160308173709.GC3017@gmail.com>
From: Dmitry Vyukov <dvyukov@google.com>
Date: Tue, 8 Mar 2016 18:41:52 +0100
Message-ID: <CACT4Y+bz-YyeYnCNQtiwqx5efkZQSm_q9q0gPfpaYjUtWpmBeQ@mail.gmail.com>
Subject: Re: [RESEND PATCH 0/5] perf core: Support overwrite ring buffer
To: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>, Wang Nan <wangnan0@huawei.com>,
        Ingo Molnar <mingo@redhat.com>, LKML <linux-kernel@vger.kernel.org>,
        He Kuang <hekuang@huawei.com>, Alexei Starovoitov <ast@kernel.org>,
        Arnaldo Carvalho de Melo <acme@redhat.com>,
        Brendan Gregg <brendan.d.gregg@gmail.com>,
        Jiri Olsa <jolsa@kernel.org>,
        Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>,
        Namhyung Kim <namhyung@kernel.org>, Zefan Li <lizefan@huawei.com>,
        pi3orama@163.com
Content-Type: text/plain; charset=UTF-8
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1668
Lines: 38

On Tue, Mar 8, 2016 at 6:37 PM, Ingo Molnar <mingo@kernel.org> wrote:
>
> * Dmitry Vyukov <dvyukov@google.com> wrote:
>
>> > fomalhaut:~/go/src/github.com/google/syzkaller> ps aux | grep -i syz
>> > mingo      1374  0.0  0.0 118476  2376 pts/2    S+   18:23   0:00 grep --color=auto -i syz
>> >
>> > and with no kernel messages in dmesg - and with a fully functional system.
>> >
>> > I'm running the 16-task load on a 120 CPU system - should I increase it to 120?
>> > Does the code expect to saturate the system?
>>
>> No, it does not expect to saturate the system. Set "procs" to 480, or
>> something like that.
>
> Does not seem to help much:
>
> fomalhaut:~> vmstat 10
> procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
>  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
>
>  1  0      0 257465904 219940 4736092    0    0     0   102 16022 4396  0  1 99  0  0
>  2  0      0 257452144 220496 4755052    0    0     2  3649 14286 4627  0  1 99  0  0
>  2  0      0 257473408 221188 4770824    0    0    15  1898 17175 4474  0  1 99  0  0
>
> Only around 1% system utilization. Should I go for 1,000 or more? :)
>
> Peter, do you experience with running syz-kaller on larger CPU count Intel
> systems?


Try to set "dropprivs": false in config.
I've noticed that creation/destruction of namespaces is very slow and
globally serialized. So sometimes it takes tens of seconds for each
worker processes to startup. For perf-related syscalls it should be
"safe" to just run as root. And perf subsystem operation is also
unaffected by namespaces as far as I know, so it should not affect
behavior as well.