Unexpected behaviour of O(1) scheduler on SMP 2.6 Kernel

• Previous message: Adrian Bunk: "Status of net/ipv4/ipvs/ip_vs_proto_icmp.c?"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Date:	Thu, 12 May 2005 23:27:46 -0500 (CDT)
To: linux-kernel@vger.kernel.org

I am running some experiments with the new O(1) scheduler. I am running my
experiments on a cluster of 10 machines, running Rocks4.0 Beta (which
comes with 2.6 kernel and hence the new scheduler)

The behaviour of scheduler is unexplained, and am getting some very
unexpected results. I read about the design of O(1) scheduler, and my
results completely DISAGREE with the theory. Any insight into explaining
this will be helpful.

--------------------------CASE 1-------------------------------------------

On a dual processor CPU, I have 2 processes running P1 and P2.

Here is snapshot of CPU utilization:

        User Sys Nice
CPU1 40 30 0 P1
CPU2 40 30 0 P2

Now I add a niced process with nice value =19. So I expect it to utilize
the left over 30% CPU on either one of the CPUs. But instead the new
snapshot of the system is

        User Sys Nice
CPU1 0 0 100 niced_process
CPU2 50 40 0 P1 & P2

So overall the nice process finishes in time. And normal process gets
slowed down by 30-40 %.

**NOTE: P1 and P2 and MPI jobs, so they wait on gettting data from other
nodes and consume only 70% CPU max.
nice process is cpu hog and consume 100% CPU
---------------------------------------------------------------------

--------------------------CASE 2-------------------------------------------

On a dual processor CPU, I have 2 processes running P1 and P2. But this
time P1 and P2 are CPU hogs themselves

Here is snapshot of CPU utilization:

        User Sys Nice
CPU1 100 0 0 P1
CPU2 100 0 0 P2

Now I add a niced process with nice value =19.

        User Sys Nice
CPU1 100 0 0 P1
CPU2 95 0 5 P2 & niced_process

This is GREAT! As expected.

But the CATCH is, if the niced_process is started first, and then P1 and
P2 are initiated, the behaviour is completely unexpected.

        User Sys Nice
CPU1 0 0 100 niced_process
CPU2 0 0 0 -

Now I add P1 and P2

        User Sys Nice
CPU1 0 0 100 niced_process
CPU2 100 0 0 P1 & P2

I expected the nice process to give up CPU to either P1 and P2.

NOTE: Here P1 , P2 and Niced_process are CPU hogs, and run at 100% cpu
utilization

---------------------------------------------------------------------

Any comments will be appreciated.

PLEASE EMAIL ME AT : dvyas@cs.uh.edu

Thanks
Deepti

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

• Previous message: Adrian Bunk: "Status of net/ipv4/ipvs/ip_vs_proto_icmp.c?"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]