rdma_listen() backlog maximum

rdma_listen() backlog maximum

[Mike Mundy]

I'm prototyping a server that needs to accept a large number of
simultaneous incoming RDMA CM connections. I noticed that
drivers/infiniband/core/ucma.c limits the backlog to 128. Is there a
reason why the limit is 128? Or to put it another way, what breaks if
the limit is increased to a larger value (say 256)?

Mike M
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: rdma_listen() backlog maximum

[Roland Dreier]

On Thu, Feb 3, 2011 at 8:40 AM, Mike Mundy <mmundy42-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> I'm prototyping a server that needs to accept a large number of
> simultaneous incoming RDMA CM connections. I noticed that
> drivers/infiniband/core/ucma.c limits the backlog to 128. Is there a
> reason why the limit is 128? Or to put it another way, what breaks if
> the limit is increased to a larger value (say 256)?

In the current kernel the backlog defaults to 1024 and there is a sysctl to
change it at runtime.

So I think you are completely safe increasing your backlog on older
kernels.  (The commit to backport is 97cb7e40c696)

 - R.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: rdma_listen() backlog maximum

[Or Gerlitz]

Mike Mundy wrote:
> I'm prototyping a server that needs to accept a large number of
> simultaneous incoming RDMA CM connections. I noticed that
> drivers/infiniband/core/ucma.c limits the backlog to 128. Is there a
> reason why the limit is 128? Or to put it another way, what breaks if
> the limit is increased to a larger value (say 256)?

My understanding is that with the current cma/cm code, under IB/IBoE
the backlog has no impact, in that environment, you're practically 
limited by the mad layer ability to receive the REQ mads. Sean, am I 
correct?

Or.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: rdma_listen() backlog maximum

[Hefty, Sean]

> My understanding is that with the current cma/cm code, under IB/IBoE
> the backlog has no impact, in that environment, you're practically
> limited by the mad layer ability to receive the REQ mads. Sean, am I
> correct?

The backlog does have an effect for user space.  At most backlog connection request events are queued in the kernel.  Additional requests that arrive while the queue is full result in the request being dropped.

- Sean
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: rdma_listen() backlog maximum

[Or Gerlitz]

Hefty, Sean <sean.hefty-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:

> The backlog does have an effect for user space.  At most backlog connection request
> events are queued in the kernel.  Additional requests that arrive while the queue is full result > in the request being dropped.

sure, I get it now.

Or.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html