Weekly status report (27th May)

[Robert N. M. Watson]

On 31 May 2011, at 14:52, George Neville-Neil wrote:

>> - Do we really need to expose queue information and threads to user
>> applications?
> 
> There are applications that will want this information.
> 
>> Probably most of BPF application requires to merge packet streams from
>> threads at last.
>> For example, sniffer app such as tcpdump and wireshark need to output
>> packet dump on a screen, before output it on the screen we need to
>> merge packet streams for each queues into one stream.
>> If so, isn't it better to merge stream in kernel, not userland?
>> 
>> 
>> I'm not really sure about use case of BPF, maybe there's use case can
>> get benefit from multithreaded BPF?
> 
> Certainly there is a case for it, but perhaps not yet.  Let's get through the
> work you've already planned first.  I see the test case isn't written yet, so
> how are you testing these changes?  When I get some time, probably next week,
> I'll want to run some of this code myself.

The rationale to for exposing queues to userspace explicitly is the same as the rationale for exposing queues to the OS: it's not just packet data that has cache issues, but program data to do with processing packet data.

The reason for having each BPF device have a input and output queue masks in my initial thinking was that we would set then to 0xffffffffffffffff by default, meaning that a particular BPF device would merge and collect all packets, picking an arbitrary ordering for interlacing the streams (as it does today). However, an application might decide to open multiple devices, one each having a particular bit set, in order to receive for a particular queue. The application could then have particular threads use particular BPF devices, and based on hardware having gotten the flow assignment right, it could then avoid cache line contention for statistics and even stateful processing of flows in its different threads. That actually gives you a spectrum between today's behaviour and greater levels of granularity, and lets the application decide "how early" to blend the different queues of data. It can ask the kernel to do it, or it can do it itself in userspace.

Robert