Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: [802.3_100GNGOPTX] Compute Performance vs Latency



Dear Task Force Members,

Please excuse my mental typo at the beginning of that message. (Task Force, Task Force, Task Force).

:)

Dan

On 1/4/13 3:56 PM, Daniel Dove wrote:
Dear Study Group Participants,

A question that has lingered in my mind relates to the additional latency of FEC as it would affect HPC and distributed computing performance in short-reach optical applications. While we can clearly understand the impact of FEC on BER, and we can assess various approaches and their impact on latency, what has been lacking is detail on the impact of latency to real-world compute applications.

I recently discovered a performance presentation at the HPC Advisory Council that is focused on comparing two particular compute platforms over a variable set of compute nodes, but what would be really interesting is to see a single compute platform and the effect strictly of channel latency on the performance. Capturing the impact of various channel latencies over the number of nodes would also be of interest.

Specifically, imagine slide 7 with a single compute platform and multiple bars (one for each latency value) per node-count. Use latency values of 10ns, 100ns, 1uS for example which could be created by using longer fibers between nodes.

I understand that different applications will have different levels of sensitivity, so using industry benchmarks seems like an appropriate way to assess the impact on performance. It would allow end users to evaluate the effect of link distances and other sources of latency such as FEC on overall compute performance.

This message is a request for contribution in this area. It would be very helpful to the Task Force to understand the tradeoffs we are making between cost, complexity, and performance. If someone has an alternative assessment that draws out the impact of link latency on distributed compute performance, please feel free to share.

Best Regards,
--
Dan Dove
Chair, P802.3bm