Re: [PATCH v2] RPS: Sparse connection optimizations - v2

[Deng-Cheng Zhu <dczhu@mips.com>]

On 05/07/2012 03:38 PM, Eric Dumazet wrote:
> On Mon, 2012-05-07 at 14:48 +0800, Deng-Cheng Zhu wrote:
>> On 05/04/2012 11:31 PM, Tom Herbert wrote:
>>>> I think the mechanisms of rps_dev_flow_table and cpu_flow (in this
>>>> patch) are different: The former works along with rps_sock_flow_table
>>>> whose CPU info is based on recvmsg by the application. But for the tests
>>>> like what I did, there's no application involved.
>>>>
>>> While rps_sock_flow_table is currently only managed by recvmsg, it
>>> still is the general mechanism that maps flows to CPUs for steering.
>>> There should be nothing preventing you from populating and managing
>>> entries in other ways.
>>
>> Well, even using rps_sock_flow_table to map the sparse flows to CPUs,
>> we still need a data structure to describe a single flow -- that's what
>> struct cpu_flow is doing. Besides, rps_sock_flow_table, by its meaning,
>> does not seem to make sense for our purpose. How about keeping the patch
>> as is but renaming struct cpu_flow to struct rps_sparse_flow? It's like:
>>
>
> sock_flow_table is about mapping a flow (by its rxhash) to cpu.
>
> If you feel 'sock' is bad name, you can rename it.
>
> You dont need adding new data structure and code in fast path.
>
> Only the first packet of a new flow might be handled by 'the wrong cpu'.
>
> If you add code in forward path to change flow_table for next packets,
> added cost in fast path is null.

Did you really read my patch and understand what I commented? When I was
talking about using rps_sparse_flow (initially cpu_flow), neither
rps_sock_flow_table nor rps_dev_flow_table is activated (number of
entries: 0).

FYI below:

On 05/04/2012 11:39 AM, Deng-Cheng Zhu wrote:
 > On 05/04/2012 11:22 AM, Tom Herbert wrote:
 >>> +struct cpu_flow {
 >>> + struct net_device *dev;
 >>> + u32 rxhash;
 >>> + unsigned long ts;
 >>> +};
 >>
 >> This seems like overkill, we already have the rps_flow_table and this
 >> used in accelerated RFS so the device can also take advantage of
 >> steering. Maybe somehow program that table for your sparse flows?
 >
 > In fact I did ever try something different in rps_flow_cnt (except for
 > rps_cpus, the only tunable thing relating to RPS in sysfs, am I
 > missing something?) and found no effect in my tests (iperf between 2
 > PCs via Malta which works as router and uses iptables/NAT+RPS)...


Deng-Cheng