Re: process starvation with 2.6 scheduler

[Mike Galbraith]

(please line wrap)

On Mon, 2006-06-05 at 12:48 -0700, Kallol Biswas wrote:
> Hello,
>        We have a process starvation problem with our 2.6.11 kernel running on a ppc-440 based system.
> 
> We have a storage SOC based on PPC-440. The SOC is emulated on a system emulator called Palladium. It is from Cadence. The system runs at 400KHz speed. It has three Ethernet ports; they are connected to outside lab network with a speed bridge.
> 
> The netperf server netserver runs on the emulated system (2.6.11 kernel on Palladium). There are netperf linux clients running on a x86 box.
> 
> If netperf request response (TCP_RR) traffic is run on all three ports; after sometime only one port remains active, the application (netperf client) on other two ports wait for a long time and eventually time out.
> 
> The netserver code has been instrumented. For one of the starved netserver processes it has been found that the TCP_RR request from the netperf client on linux x86 box has been received by the server, it has issued send() call to send back reply but send() never returns.
> 
> With an ICE connected to the Palladium (emulator) I have dumped the kernel data structures of the starved process and the active process. 
> 
> 
> For Active  Process:
>   Time_slice 84
>   Policy : SCHED_NORMAL
>   Dynamic priority: 118
>   Static priority: 120
>   Preempt_count: 0x20100
>   Flags = 0
>   State = 0 (TASK_RUNNING)
> 
> For Starved Process:
>   Time slice: 77
>   Policy: SCHED_NORMAL
>   Dynamic priority: 120
>   Static priority: 120
>   Preempt_count: 0x10000000 (PREEMPT_ACTIVE is set)
>   Flags = 0 
>   State = 0 (TASK_RUNNING)
> 
> Any help to debug the problem is welcome. 

I'm having difficulty understanding.  Are you saying that the "starved"
tasks are runnable, but receiving _zero_ cpu?  That's impossible with
only one other SCHED_NORMAL task afaik, which makes me think you may
mean they're not receiving cpu frequently enough to keep clients from
timing out?  One task which has slept enough to acquire interactive
status (as above) can hold others off the cpu for quite a while if it
starts a burst of heavy cpu burning.  If your netperf clients are
choking on this latency, running the servers at nice 19 should prevent
the problem.

	-Mike

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/