Abstract
Datacenter servers are stepping into an era marked by powerful multi- /many-core processors. Severe problems such as I/O contentions in those large-scale platforms pose an unprecedented challenge. Prior studies primarily considered I/O bandwidth as a major performance bottleneck. However, ourwork reveals that in many cases the fundamental cause of I/O contentions is the inefficiency of OS schedulers. Particularly, the modern system is not aware of this fact and thus suffers from poor I/O performance, especially for datacenter servers. Based on our findings, we propose a new software-based scheduling approach, WiseThrottling, to reduce I/O contention. WiseThrottling performs asynchronous and self-adjustment scheduling for concurrent tasks.We evaluate our approach across a wide range of C/OpenMP/MapReduce workloads on a 64-core server in Dawning Cluster datacenter. The experimental results exhibit that WiseThrottling is effective for reducing the I/O bottleneck and it can improve the overall system performance by up to 207 %.
Original language | English (US) |
---|---|
Article number | A013 |
Pages (from-to) | 3054-3093 |
Number of pages | 40 |
Journal | Journal of Supercomputing |
Volume | 71 |
Issue number | 8 |
DOIs | |
State | Published - Aug 1 2015 |
Bibliographical note
Funding Information:This research is supported by the National High Technology Research and Development Program of China under Grants No. 2012AA010902 and 2015AA011505; the NSFC under Grants No. 61202055, 61221062, 61303053, 61432016 and 61402445; and the National Basic Research Program of China under Grant No. 2011CB302504.
Publisher Copyright:
© Springer Science+Business Media New York 2015.
Keywords
- I/O contention
- Multi-/many-core server
- Resource description
- Scheduling