As scientists and researchers, we take the reproducibility of our work very seriously. We don't expect you to trust our results, in fact, we hope that you don't! On this page, we provide links to detailed descriptions of the experiments behind each of the figures in our research publications. The full experimental descriptions include the precise configuration of our test equipment, links to the source code for our tools, patches that we made to other peoples' tools and original preprocessed data-sets that we gathered. Our hope is that anyone can use these descriptions to re-run any of our experiments. We believe that this kind of openness is the way that all good, scholarly scientific research should be conducted.

For details of our publications, including links to the manuscripts, please see the publications page.

NSDI 2015 – Queues don't matter when you can jump them!

Figure Description
Fig. 1a PTPd synchronization offset with and without sharing the network with Hadoop MapReduce.
Fig. 1b Memcached request latencies with and without sharing the network with Hadoop MapReduce.
Fig. 1c Naiad barrier synchronization latencies with and without sharing the network with Hadoop MapReduce.
Tbl. 1 Latencies observed as ping and iperf share various parts of the network.
Fig. 3a Ping packet latency across a switch with and without QJump enabled.
Fig. 3b QJump reduced memcached request latency variance in the presence of Hadoop MapReduce traffic.
Fig. 3c QJump fixes Naiad barrier synchronization latency in the presence of Hadoop MapReduce traffic.
Fig. 5 PTPd, memcached and Hadoop sharing a cluster, with and without QJump enabled.
Fig. 6 QJump offers constant two-phase commit throughput even at high levels of network interference.
Fig. 7 QJump comes closest to ideal performance when compared with Ethernet Flow Control, ECN and DCTCP.
Fig. 9 Normalized flow completion times in a 144-host simulation. QJump outperforms stand-alone TCP, DCTCP and pFabric for small flows.
Fig. 10 Memcached throughput and latency as a function of the QJump rate limits.
Fig. 11 Latency bound validation of QJump with 60 host generating full rate, fan-in traffic.

;login: April 2015 – Jump the queue to lower latency

Figure Description
Fig. 1a PTPd synchronization offset with and without sharing the network with Hadoop MapReduce.
Fig. 1b Memcached request latencies with and without sharing the network with Hadoop MapReduce.
Fig. 1c Naiad barrier synchronization latencies with and without sharing the network with Hadoop MapReduce.
Fig. 3 QJump comes closest to ideal performance when compared with Ethernet Flow Control, ECN and DCTCP.