[ovs-discuss] cycles_count_xxx CPU overhead

Daniele Di Proietto diproiettod at ovn.org
Mon Jan 16 04:31:49 UTC 2017


2017-01-12 23:47 GMT-08:00 nickcooper-zhangtonghao <nic at opencloud.tech>:
> The dp_netdev_process_rxq_port() will use the “cycles_count_start” and
> “cycles_count_end” for counting the cycles.
> But via perf tool, I found that cycles_count_xxx uses a lot of CPU. One
> question to ask: there is one thread to use the
> pmd struct to processing the packets, right? If so, pmd should be thread
> safety and atomic should not be used for
> cycles_count_xxx.

'pmd->cycles' is also read by the main thread.

The idea is that we don't care about being too accurate, but we
shouldn't be totally inconsistent.

As explained in 'non_atomic_ullong_add()', we're not doing atomic
increments.  We're also not using any memory barrier ('relaxed').

By using atomic operation we're only asking the compiler to read/write
the 64-bit word in a single operation, to prevent reading the most
significant 32-bits from an old version and the least significant from
a new version.

This should have no impact on 64-bit architectures (the generated asm
should be the same as the non atomic case), but might have an effect
on 32-bit architectures.

I suspect that the functions might show up in perf only when the
thread is not processing packets, because we're executing very few
lines of code over and over.

Does this make sense to you?

Thanks,

Daniele

>
>
> Thanks.
> Nick
>
>
> _______________________________________________
> discuss mailing list
> discuss at openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
>


More information about the discuss mailing list