The Case for Tiny Tasks in Compute Clusters

TitleThe Case for Tiny Tasks in Compute Clusters
Publication TypeConference Paper
Year of Publication2013
AuthorsOusterhout, K., Panda A., Rosen J., Venkataraman S., Xin R., Ratnasamy S., Shenker S., & Stoica I.
Page(s)1-6
Other Numbers3435
Abstract

We argue for breaking data-parallel jobs in compute clustersinto tiny tasks that each complete in hundreds of milliseconds.Tiny tasks avoid the need for complex skewmitigation techniques: by breaking a large job into millionsof tiny tasks, work will be evenly spread over availableresources by the scheduler. Furthermore, tiny tasksalleviate long wait times seen in today’s clusters for interactivejobs: even large batch jobs can be split into smalltasks that finish quickly. We demonstrate a 5.2x improvementin response times due to the use of smaller tasks.In current data-parallel computing frameworks, hightask launch overheads and scalability limitations preventusers from running short tasks. Recent research has addressedmany of these bottlenecks; we discuss remainingchallenges and propose a task execution framework thatcan efficiently support tiny tasks.

Acknowledgment

We thank Matei Zaharia, Colin Scott, John Ousterhout,and Patrick Wendell for useful feedback on earlierdrafts of this paper. This research is supported inpart by NSF CISE Expeditions award CCF-1139158 andDARPA XData Award FA8750-12-2-0331; gifts fromAmazon Web Services, Google, SAP, Blue Goji, Cisco,Clearstory Data, Cloudera, Ericsson, Facebook, GeneralElectric, Hortonworks, Huawei, Intel, Microsoft, NetApp,Oracle, Quanta, Samsung, Splunk, VMware andYahoo!; and a Hertz Foundation Fellowship.

URLhttp://www.icsi.berkeley.edu/pubs/networking/ICSI_thecasefortiny13.pdf
Bibliographic Notes

Proceedings of the 14th USENIX Workshop on Hot Topics in Operating Systems, pp. 1-6, Santa Ana Pueblo, New Mexico

Abbreviated Authors

K. Ousterhout, A. Panda, J. Rosen, S. Venkataraman, R. Xin, S. Ratnasamy, S. Shenker and I. Stoica

ICSI Research Group

Networking and Security

ICSI Publication Type

Article in conference proceedings