"Dan Shimshoni" <[EMAIL PROTECTED]> writes:
> So the question is:
> On a principle level, suppose we will want to use RDMA instead for
> TCP between A and B. (with 1Gb). How much can using RDMA improve
> performance ? Does it worth it ? or should we think of infiniband
> solutions like Meallnox HBAs?
InfiniBand will give you _much_ better bandwidth and latency for
sure. This does not necessarily mean you will speed up your
application by a factor of 20 (20 Gbps bandwidth is common nowadays
for IB): the ultimate gain will strongly depend on workload.
> RDMA (when using 1Gb and infiniband)? is it a lot of work ?
> do you know a good resource for this task (I saw some tutorial
> about RDMA API written by Roland Drier).
If you are interested in InfiniBand, check out things like IPoIB (IP
over InfiniBand), SDP (Socket Direct Protocol), etc. Just google for
starters. The effort may be not large at all.
Basically, SDP "hijacks" stream sockets (SOCK_STREAM) and allows
bypassing of the OS TCP/IP stack ("kernel bypass") and utilizes
various InfiniBand features (including RDMA) for high performance data
transfer. Other socket types will use IPoIB stack. SDP also depends on
IPoIB drivers for IP address resolution, local address assignment,
etc. This can be done transparently for the application above the
socket layer (in the simplest case, try invoking your app with
LD_PRELOAD=/usr/lib/libsdp.so or whatever is appropriate).
Disclaimers:
1) YMMV.
2) I work for an InfiniBand company.
--
Oleg Goldshmidt | [EMAIL PROTECTED] | http://www.goldshmidt.org
=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]
