Researchers boost multi-core CPU performance with better prefetching

Piling on cores is one way to boost performance, but it's not necessarily the most efficient way -- researchers at North Carolina State University have developed a new prefetching technique for processors that could boost performance by up to 40-percent. As you may know, any data not stored in a CPU's cache must be pulled from RAM, but as more cores are added they can create a bottleneck by competing for memory access. To counter this designers use prefetching to predict what information will be needed and grab it ahead of time, but guessing wrong can hurt performance. Researchers tackled this problem from two fronts: first, by creating a better algorithm for divvying up bandwidth, and second, by selectively turning off prefetching when it might slow the CPU. Full PR and an abstract of the study being published June 9th are after the break.

Show full PR text

New Bandwidth Management Techniques Boost Operating Efficiency In Multi-Core Chips
For Immediate Release

Release Date: 05.25.2011
Filed under Releases

Researchers from North Carolina State University have developed two new techniques to help maximize the performance of multi-core computer chips by allowing them to retrieve data more efficiently, which boosts chip performance by 10 to 40 percent.

To do this, the new techniques allow multi-core chips to deal with two things more efficiently: allocating bandwidth and "prefetching" data.

Multi-core chips are supposed to make our computers run faster. Each core on a chip is its own central processing unit, or computer brain. However, there are things that can slow these cores. For example, each core needs to retrieve data from memory that is not stored on its chip. There is a limited pathway – or bandwidth – these cores can use to retrieve that off-chip data. As chips have incorporated more and more cores, the bandwidth has become increasingly congested – slowing down system performance.

One of the ways to expedite core performance is called prefetching. Each chip has its own small memory component, called a cache. In prefetching, the cache predicts what data a core will need in the future and retrieves that data from off-chip memory before the core needs it. Ideally, this improves the core's performance. But, if the cache's prediction is inaccurate, it unnecessarily clogs the bandwidth while retrieving the wrong data. This actually slows the chip's overall performance.

"The first technique relies on criteria we developed to determine how much bandwidth should be allotted to each core on a chip," says Dr. Yan Solihin, associate professor of electrical and computer engineering at NC State and co-author of a paper describing the research. Some cores require more off-chip data than others. The researchers use easily-collected data from the hardware counters on each chip to determine which cores need more bandwidth. "By better distributing the bandwidth to the appropriate cores, the criteria are able to maximize system performance," Solihin says.

"The second technique relies on a set of criteria we developed for determining when prefetching will boost performance and should be utilized," Solihin says, "as well as when prefetching would slow things down and should be avoided." These criteria also use data from each chip's hardware counters. The prefetching criteria would allow manufacturers to make multi-core chips that operate more efficiently, because each of the individual cores would automatically turn prefetching on or off as needed.

Utilizing both sets of criteria, the researchers were able to boost multi-core chip performance by 40 percent, compared to multi-core chips that do not prefetch data, and by 10 percent over multi-core chips that always prefetch data.

The paper, "Studying the Impact of Hardware Prefetching and Bandwidth Partitioning in Chip-Multiprocessors," will be presented June 9 at the International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS) in San Jose, Calif. The paper was co-authored by Dr. Fang Liu, a former Ph.D. student at NC State. The research was supported, in part, by the National Science Foundation.

NC State's Department of Electrical and Computer Engineering is part of the university's College of Engineering.

-shipman-

Note to Editors: The study abstract follows.

"Studying the Impact of Hardware Prefetching and Bandwidth Partitioning in Chip-Multiprocessors"

Authors: Fang Liu and Yan Solihin, North Carolina State University

Presented: June 9, 2011, at the International Conference on Measurement and Modeling of Computer Systems, San Jose, Calif.

Abstract: Modern high performance microprocessors widely employ hardware prefetching to hide long memory access latency. While useful, hardware prefetching tends to aggravate the bandwidth wall, a problem where system performance is increasingly limited by the availability of off-chip pin bandwidth in Chip Multi-Processors (CMPs). In this paper, we propose an analytical model-based study to investigate how hardware prefetching and memory bandwidth partitioning impact CMP system performance and how they interact. The model includes a composite prefetching metric that can help determine under which conditions prefetching can improve system performance, a bandwidth partitioning model that takes into account prefetching effects, and a derivation of the weighted speedup-optimum bandwidth partition sizes for different cores. Through model-driven case studies, we find several interesting observations that can be valuable for future CMP system design and optimization. We also explore simulation-based empirical evaluation to validate the observations and show that maximum system performance can be achieved by selective prefetching, guided by the composite prefetching metric, coupled with dynamic bandwidth partitioning.