On Sat, Feb 28, 2009 at 2:36 PM, Andi Kleen <andi@firstfloor.org> wrote:
quoted text
>> The previous solution for IPC that Oracle was using was based on UDP,
>> which I think could be considered very close to using raw sockets --
>> each process is responsible for its own acks, retransmits, everything.
>> Doing this on a highly loaded machine resulted in a cascade where
>> performance got worse and worse.
>
> Could you describe that cascade in more detail?
> The problem was that the retransmits didn't have high enough priority?

I think the gist of it is:

Higher load -> more time before a process runs -> rcvbuf overfills ->
ACKs dropped -> timeouts -> more retransmissions -> even higher load.

Things are fine until they hit a point where everything goes to hell.

quoted text
>> Additionally, our interconnect is primarily Infiniband. It natively
>> implements a reliable datagram connection type so RDS leverages that.
> So perhaps it would make more sense to have a thin direct interface
> to that IB service? Or perhaps it already exists? (I admit I don't know
> the IB interfaces very well)

The most direct userspace API is uDAPL -- apps can create IB
connections (queue pairs) directly. This was tried but didn't work out
so well. A queue pair (QP) is a TX/RX ring -- a nontrivial amount of
memory. If each process needs a new QP to talk to every other process
then the number of RAM-hungry QPs becomes huge.

RDS is only slightly less direct -- apps don't create queue pairs,
they create RDS sockets. RDS uses only one QP for all traffic to each
remote node, so the number of QPs on a node is equal to the number of
remote nodes, as opposed to (number of local processes * number of
remote processes).

Regards -- Andy
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html