Re: [BUG] soft lockup while booting machine with more than 700 cores

[Dimitri Sivanich <sivanich@sgi.com>]

On Thu, Feb 10, 2011 at 03:12:23PM -0600, Jack Steiner wrote:
> On Thu, Feb 10, 2011 at 01:03:25PM -0800, David Miller wrote:
> > From: Jack Steiner <steiner@sgi.com>
> > Date: Thu, 10 Feb 2011 14:56:48 -0600
> > 
> > > We also noticed that the rebalance_domains() code references many per-cpu
> > > run queue structures. All of the structures have identical offsets relative
> > > to the size of a cache leaf. The result is that all index into the same lines in the
> > > L3 caches. That causes many evictions. We tried an experimental to
> > > stride the run queues at 128 byte offsets. That helped in some cases but the
> > > results were mixed.  We are still experimenting with the patch.
> > 
> > I think chasing after cache alignment issues misses the point entirely.
> > 
> > The core issue is that rebalance_domains() is insanely expensive, by
> > design.  It's complexity is N factorial for the idle non-HZ cpu that is
> > selected to balance every single domain.
> > 
> > A statistic datastructure that is approximately 128 bytes in size is
> > repopulated N! times each time this global rebalance thing runs.
> > 
> > I've been seeing rebalance_domains() in my perf top output on 128 cpu
> > machines for several years now.  Even on an otherwise idle machine,
> > the system churns in thus code path endlessly.
> 
> Completely agree! Idle rebalancing is also a big problem. We've seen
> significant improvements on large systems in network thruput by
> disabling IDLE load balancing for the higher (2 & 3) scheduling domains.
> 
> This is not a real fix but points to a problem.
>

Here are some TCP STREAMS test numbers from a large, otherwise idle UV system.

With SD_BALANCE_NEWIDLE turned on for all domain levels:

TCP STREAM TEST from localhost (::1) port 0 AF_INET6 to localhost (::1) port 0
AF_INET6 : cpu bind
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec
 87380  16384  16384    10.00     115.32

With SD_BALANCE_NEWIDLE turned off for domain levels 2 & 3 (NODES & ALLNODES):
 87380  16384  16384    10.00    14685.51

I am curious as to why there would be such a large discrepancy.