[Spread-users] Spread Accelerated Ring Experimental version 4.3.0 released
yairamir at cs.jhu.edu
Tue Jul 9 01:42:29 EDT 2013
I wanted to follow up on Amy's message and encourage people to give this
experimental version of Spread a try.
On 1 Gig networks, this version is practically maxing the hardware, able to achieve
performance close to the theoretical hardware limit in demanding benchmarks with many senders
and receivers (over 900Mbits/sec sustained total ordered multicast with virtual synchrony semantics).
On 10 Gig networks, this version is doing considerably higher than any reported protocol in the field
I am aware of.
We hope to make a version based on this protocol the next release of Spread, so any comment
good, bad, or neutral will be appreciated. (getting 30%-50% throughput increase while at the same
time getting %20-35% latency reduction seems appealing to us).
:) Yair. www.cs.jhu.edu/~yairamir
On 7/3/13 5:22 PM, Amy Babay wrote:
> The Spread Toolkit Accelerated Ring Experimental release 4.3.0 has been released. This version features an experimental protocol and was developed as part of research being done at the Distributed Systems and Networks Lab at Johns Hopkins University. You read about and download it here: http://www.spread.org/AcceleratedRingExperimentalRelease.html.
> The new protocol is tailored for data center networks, and my experiments show that it can provide a 30-50% increase in throughput and a 20-35% reduction in latency compared to the original protocol in local area networks. We would encourage you to test it out and would love to hear about your experiences and any issues you encounter.
> For those interested in performance-tuning for the Accelerated Ring protocol, the relevant configuration parameters are AcceleratedRing and AcceleratedWindow. AcceleratedRing must be set to "true" in the spread.conf file to enable the Accelerated Ring protocol. AcceleratedWindow specifies the number of packets from the PersonalWindow that will be sent after the token. In general, the higher (closer to the full PersonalWindow) the AcceleratedWindow is, the faster the token circulates, and the greater the throughput and latency benefits are. However, for large values of PersonalWindow (e.g. 100), setting AcceleratedWindow too high can lead to reduced throughput due to retransmissions.
> The sample.spread.conf file included in the release includes settings of AcceleratedRing=true, AcceleratedWindow=20, PersonalWindow=20, and GlobalWindow=100, which we found to provide a good performance tradeoff in our experiments for 1-Gig and 10-Gig networks.
> Amy Babay
> Spread-users mailing list
> Spread-users at lists.spread.org
More information about the Spread-users