- Our work on configuration tuning for distributed databases in the face of changing workloads has been accepted to Usenix ATC. Acceptance rate was 71/356 = 19.9%. [ PDF ]
This work shows how you can reconfigure a distributed system in the face of changing workloads, taking into account the cost and the benefit of reconfiguration. One of the evaluation applications is the workload of the metagenomics portal run by our NIH collaborator, Folker Meyer at Argonne National Lab. We show that extensions to previous configuration strategies that do not take cost-benefit of reconfiguration into account degrades the throughput below the default settings. We demonstrate the power of our approach on three real-world application traces.
- Congratulations to students Ashraf and Alex (Purdue) and collaborators, Paul (JHU, work done while at Purdue), Subrata (Adobe Research), Folker (Argonne National Lab), and Somali (Purdue).
- Our work with Department of Energy/Lawrence Livermore National Lab (DOE/LLNL) on approximating scientific computation through reducing precision of floating point variables gets accepted into ICS (International Conference on Supercomputing). Acceptance rate was 45/193 = 23.3%.
The work shows how you can quickly find which variables to reduce precision of and to what extent, while bounding the accuracy loss. Our approach, unlike prior works in the area which work on benchmark-style program, is shown to apply to a realistic large-scale application, a widely studied proxy application in DOE co-design efforts for exascale called Lulesh. [ PDF ]
- Congratulations to Pradeep (Purdue), Ranvijay (NVIDIA), Paul (JHU), and Ignacio (DOE). Ranvijay and Paul did this work while they were here at Purdue.