[OmniOS-discuss] Ang: Re: Ang: Re: infiniband

[Narayan Desai]

On Fri, Oct 24, 2014 at 7:55 AM, Michael Rasmussen <mir at miras.org> wrote:

>
> I think the issue here is that OpenSM is a single threaded application
> so when the CPU which is assigned to it maxes out performance will
> drop. So I guess it comes down to a trade-off between the
> performance of the host CPU and the performance of the switch hardware.
>

You're right in a general sort of sense; if the opensm process is resource
starved, that will result in slower performance for subnet manager
operations. That said, the subnet manager isn't involved in data transfers,
except to program switch forwarding tables and answer path requests.
Neither of these are particularly important for data transfers, unless the
topology is constantly changing (like if links are flapping), or the
network is otherwise sick. I wouldn't expect subnet manager performance to
effect infiniband transfer performance, unless it was picking poor routes.
That would be a function of opensm routing algorithm choice/implementation,
not of opensm performance.

I'd be much more concerned about fixing bugs in the subnet manager, which
is trickier in the case of switch-based subnet managers. I've seen all
sorts of weird subnet manager failures over the years, including the subnet
manager completely going out to lunch, which can be fixed by upgrading to
newer versions.
 -nld
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20141024/d5655bee/attachment.html>