Re: [MirageOS-devel] Parallelizing writing to network devices

[Thomas Leonard <talex5@xxxxxxxxx>]

On 17 December 2014 at 18:05, Masoud Koleini
<masoud.koleini@xxxxxxxxxxxxxxxx> wrote:
> Thanks Thomas for the great tracing tool!
>
> The following is a very simple unikernel with two interfaces, which
> redirects frames captured on the first interface to the second one:
>
> https://github.com/koleini/parallelisation
>
> The problem is that in a high packet rate (more than 80'000 pps), switch
> stops receiving. The goal is to spot the problem and enhance the throughput
> of Mirage netif.
>
> Test environment consists of another vm running a traffic generator and
> sending frames of a specific pattern (UDP frames of size 100 bytes) over the
> bridge that connects to the first interface of the unikernel. Unikernel
> forwards frames by collecting a number of frames from input queue and
> running the same number of threads that write them to the output interface.
>
> Two trace files are uploaded to the repo. The first file is the output of
> this configuration. This trace shows that each netif write locks until the
> thread that writes on the front-end connection to the ring is returned
> (function write_already_locked.)

Do these traces show it after it stopped? The second has a long sleep,
while the first looks like it was in the middle of a run.

If it had stopped in both cases, it suggests that the whole unikernel
stopped (not just the listen thread), because there are no more timer
interrupts and no sleep region.

Does "xl top" show the unikernel still using the CPU? Or it is
waiting, or crashed?

If you have a thread writing a string to the console once per second,
does it continue after unikernel stops accepting frames?

> For the second trace, the return of the thread is ignored (commenting out
> "lwt () = th in" in write_already_locked). This considerably increases
> switching speed, but after some running time, it looks that after garbage
> collection, similar problem happens.
>
> Thomas and Anil, any idea from given traces, and how it is possible to make
> the traces more informative?
>
> Thanks.
>
>
> On 28/11/14 16:55, Thomas Leonard wrote:
>>
>> On 28 November 2014 at 16:24, Anil Madhavapeddy <anil@xxxxxxxxxx> wrote:
>>>>
>>>> On 28 Nov 2014, at 16:03, Masoud Koleini
>>>> <masoud.koleini@xxxxxxxxxxxxxxxx> wrote:
>>>>
>>>> Thanks Anil.
>>>>
>>>>> - graph the ring utilisation to see if it's always full (Thomas
>>>>> Leonard's profiling patches should help here)
>>>>
>>>> Would you please point me out to the profiling patches?
>>>
>>> See:
>>> http://roscidus.com/blog/blog/2014/10/27/visualising-an-asynchronous-monad/
>>
>> The installation instructions here are for the previous version
>> (though they should still work). If you want to try the latest
>> version, the current Git mirage allows you to pass a ~tracing argument
>> to "register" in your config.ml, e.g.
>>
>> let tracing = mprof_trace ~size:1000000 () in
>> register "myunikernel" ~tracing [
>>    main $ ...
>> ]
>>
>> This uses a newer version of the profiling API. You should generally
>> "opam pin" the #tracing2 branches rather than #tracing to use it.
>>
>> Note also that it doesn't currently record ring utilisation, so you'll
>> still need to do some work to get that. You could use the
>> MProf.Counter interface, in which case the GUI will display it as a
>> graph over the trace.
>>
>>>>> - try to reduce the parallelisation to see if some condition there
>>>>> alleviates the issue to track it down.
>>>>
>>>> Reducing the maximum number of threads running in parallel reduced CPU
>>>> utilization, and vm was functioning for a much longer time, but the same
>>>> problem occurred at the end.
>>>>
>>>> It might be more useful looking at the code. Please have a look at the
>>>> function "f_thread" in the file uploaded on the following repo:
>>>>
>>>> https://github.com/koleini/parallelisation
>>>
>>> That's a lot of code to try and distill down a test case.  Try to cut it
>>> down significantly by building a minimal Ethernet traffic generator that
>>> outputs frames with a predictable pattern in the frame, and a receiver that
>>> will check that the pattern is received as expected.
>>>
>>> Then you can try out your parallel algorithm variations on the simple
>>> Ethernet sender/receiver and narrow down the problem without all the other
>>> concerns.
>>>
>>> Once the bug is tracked down, we can add the sender/receiver into
>>> mirage-skeleton and use it as a test case to ensure that this functional
>>> never regresses in the future.  Line rate Ethernet transmission has worked
>>> in the past, but we never added a test case to ensure it stays working.
>>>
>>> Anil
>>> _______________________________________________
>>> MirageOS-devel mailing list
>>> MirageOS-devel@xxxxxxxxxxxxxxxxxxxx
>>> http://lists.xenproject.org/cgi-bin/mailman/listinfo/mirageos-devel
>>
>>
>>
>
>
>
>
>
> This message and any attachment are intended solely for the addressee and
> may contain confidential information. If you have received this message in
> error, please send it back to me, and immediately delete it.   Please do not
> use, copy or disclose the information contained in this message or in any
> attachment.  Any views or opinions expressed by the author of this email do
> not necessarily reflect the views of the University of Nottingham.
>
> This message has been checked for viruses but the contents of an attachment
> may still contain software viruses which could damage your computer system,
> you are advised to perform your own checks. Email communications with the
> University of Nottingham may be monitored as permitted by UK legislation.
>



-- 
Dr Thomas Leonard        http://0install.net/
GPG: 9242 9807 C985 3C07 44A6  8B9A AE07 8280 59A5 3CC1
GPG: DA98 25AE CAD0 8975 7CDA  BD8E 0713 3F96 CA74 D8BA

_______________________________________________
MirageOS-devel mailing list
MirageOS-devel@xxxxxxxxxxxxxxxxxxxx
http://lists.xenproject.org/cgi-bin/mailman/listinfo/mirageos-devel