Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
PerfBound: conserving energy with bounded overheads in on/off-based HPC interconnects
Saravanan K., Carpenter P. IEEE Transactions on Computers67 (7):960-974,2018.Type:Article
Date Reviewed: Sep 26 2018

Network performance is crucial in high-performance computing (HPC), where individual nodes can work together to solve a large problem. Traditionally, network devices are always on, so as not to cause a message delivery delay. As more network devices are involved in HPC, an energy saving issue emerges since network devices consume a significant amount of energy. This paper focuses on how to reduce energy consumption while keeping performance acceptably high.

The main idea of this paper is a stall-timer. The stall-timer mechanism determines when the device goes to sleep after processing a frame. If the following frame comes quickly, it would be better not to go into sleep mode and thus avoid an activation delay. A key challenge is how to bound the number of messages suffering from activation delay due to arriving after a stall-timer expiration. Especially in HPC, since messages can travel along different numbers of hops, a network device will process messages with different cumulative activation delays. This paper proposes PerfBound, a method that determines the stall-timer using a histogram of idle periods so that each device can maintain a local performance bound that composes a given global performance bound (for example, one percent).

This method is effective, but it is dependent on an application. To extend the idea for multiple applications, the paper uses both the deep-sleep mode and the shallow-sleep mode of network devices. Whereas the deep-sleep mode saves significant energy but requires a long activation delay, the shallow-sleep mode consumes more energy than the deep-sleep mode but requires a much shorter activation delay. By exploiting two stall-timers for the two sleep modes, the stall-timer mechanism is applicable to various applications. Performance delays and energy savings vary depending on the application, but overall the proposed mechanism saves 40 to 70 percent energy with performance overhead bounded to only one percent.

Currently, interconnected links consume about 7.2 percent of the system energy. As more HPC is deployed and the interconnection scale increases, the issue of network energy savings will become more important. In addition to the energy savings effect with bounded performance, the proposed mechanism has several advantages. One is that it does not require application modification, which makes the mechanism deployable to existing sites. Another is that it does not require any additional communication overhead since it uses only local information.

Reviewer:  Seon Yeong Han Review #: CR146251 (1812-0638)
Bookmark and Share
 
Local and Wide-Area Networks (C.2.5 )
 
Would you recommend this review?
yes
no
Other reviews under "Local and Wide-Area Networks": Date
Microcomputer LANs (2nd ed.)
Hordeski M., TAB Books, Blue Ridge Summit, PA, 1991. Type: Book (9780830634248)
Jul 1 1992
High-speed local area networks and their performance
Abeysundara B., Kamal A. ACM Computing Surveys 23(2): 221-264, 1991. Type: Article
Jun 1 1992
Local area networking
Naugle M., McGraw-Hill, Inc., New York, NY, 1991. Type: Book (9780070464551)
Jun 1 1992
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy