[Ndn-interest] Multi-path retrieval performance
nicholas.h.briggs at gmail.com
Mon Sep 18 14:32:36 PDT 2017
Including whether these are real or virtual machines and whether the underlying physical transports are independent.
Also need to consider the interest lifetimes -- increase the pipeline too much and the interests can time out before they're responded to, requiring retransmits.
I ran the same kind of testing with a number of virtual machines on Amazon, and on some fairly beefy 10 Gb/s connected systems, testing the interest forwarding strategies when we first implemented them in ccnx 0.8.x; it's more sensitive to the conditions than you might hope, but as you are suggesting, if you have a complete top-to-bottom understanding of the setup it is possible to explain the performance one sees.
> On Sep 18, 2017, at 2:22 PM, Klaus Schneider <klaus at cs.arizona.edu> wrote:
> + What's the exact simulation setup? Topology, link bandwidth, delays, etc. ?
> On 09/17/2017 03:08 PM, Mohammad Alhowaidi wrote:
>> I was trying to do a load balancing (multi-path retrieval) experiments to get a fast retrieval for a file.
>> for example I have a router connected to three producers. and I used ndncatchunk to retrieve a file (1GB) and I started to increase the interest pipeline value.
>> I found that when the value of interest-pipeline is 10 is the best when retrieving from one producer (not sure *WHY*? ), then I tried to use 2 producers with interest-pipeline is 20 and get a better performance (retrieving chunks was divided evenly between these two producers).
>> But when I used 3 producers the performance reduced, whatever I choose the interest-pipeline value. Shouldn't retrieving the file from 3 producers with 30 interest-pipeline value be better? what could be the problem?
>> Sorry for the long question!
>> Ndn-interest mailing list
>> Ndn-interest at lists.cs.ucla.edu
> Ndn-interest mailing list
> Ndn-interest at lists.cs.ucla.edu
More information about the Ndn-interest