***SPAM*** Re: [ofa-general] [PATCH 0/26] Reliable Datagram Sockets (RDS), take 2

[Andrew Grover]

On Fri, Feb 27, 2009 at 9:56 PM, Andi Kleen <andi at firstfloor.org> wrote:
> On Fri, Feb 27, 2009 at 05:53:19PM -0800, Andrew Grover wrote:
>> On Fri, Feb 27, 2009 at 9:08 AM, Andi Kleen <andi at firstfloor.org> wrote:
>> >> This patchset against net-next adds support for RDS sockets. RDS is an
>> >> Oracle-originated protocol used to send IPC datagrams (up to 1MB)
>> >> reliably, and is used currently in Oracle RAC and Exadata products.
>> >
>> > Perhaps I missed it earlier, but what is the rationale for putting
>> > this as a socket type into the kernel? I assume they also work
>> > directly as implemented in user space using raw sockets or similar,
>> > don't they?
>>
>> You want me to implement my fancy protocol in userspace???
>
> I just asked why you're putting it in kernel space.
>
>> Do I even get to write it in C or do I need to use Ruby?
>
> Well normally people who add new subsystems to the kernel explain
> why they do that. Perhaps it's obvious to you, but at least to
> me it isn't.

Sure thing, sorry to be flippant :-)

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. Moving this to kernel code made a big
difference.

Additionally, our interconnect is primarily Infiniband. It natively
implements a reliable datagram connection type so RDS leverages that.
RDS multiplexes all processes' traffic between two hosts over a single
IB connection. Since RDS is managing IB connections at the host level
(but based on socket traffic) this is also more naturally a fit for
kernel code.

Regards -- Andy