Finding the Door in the Memory Wall, Part 2

By Erik Hagersten

March 23, 2009

Trading Parallelism for Performance

It is a common belief that only sequential applications need to be adapted for parallel execution on multicore processors. However, many existing parallel algorithms are also a poor fit. They have simply been optimized for the wrong design parameters.

In the past we have been striving for algorithms to maximize parallelism and at the same time minimize the communication between the threads. For multicore processors, however, the cost of thread communication is relatively cheap as long as the communicated data resides in a cache shared by the threads. Also, the amount of parallelism that can be explored by a multicore processor is limited by its number of cores multiplied by the number of threads running on each core. Instead, a third parameter is gaining importance for parallel multicore applications: the memory usage.

In this, the second article of the series, we contrast the behavior of a highly parallel state-of-the-art algorithm with that of a moderately-parallel algorithm in which some of the parallelism has been traded for lower DRAM bandwidth demands. We show the latter outperforms the highly parallel algorithm by a factor three on today’s multicore processors. The techniques used and some of the performance numbers are summarized here. A more detailed description of the algorithms discussed in this paper was presented at ICS 2006 together with colleagues and students from Uppsala University.

Highly Parallel Algorithm

The Gauss-Seidel algorithm (GS) is used to smooth an array with NxN elements. The original GS algorithm is pictured in Figure 1a. The new value (yellow) for each element of an NxN array is calculated as the average of its own and its four neighbors’ values. The elements of the array are updated row-wise. The element numbers in the figure refer to their iteration age. At the end of each iteration, convergence is checked and, if the condition is not met, the array will be iterated again. Typically, the array is iterated 10–30 times before the convergence is met. The red arrows in Figure 1a indicate the data dependences of this algorithm. The new values to the left of and above the yellow element have to be calculated before the yellow value can be calculated. These data dependencies make the original algorithm hard to parallelize. Memory Wall Part 2 - Fig 1

Figure 1b shows the popular red/black variation of the algorithm, where only every other element is updated in a sweep of the array (the update of red elements is shown in Figure 1b). In a second sweep, the other (black) elements are updated. Unlike the original scheme, this red/black algorithm has no data dependencies during sweeps since red elements do not depend on any other red elements. In other words, all the elements of a sweep can theoretically be updated in parallel – its parallelism is N2/2.

Figure 1c shows how two cores may divide the work. This scheme keeps the communication between the cores at a minimum: only values of the element on the boarder between the threads need to be communicated, and the threads only need to synchronize once per sweep. So, according to the old definition of a good algorithm, the red/black algorithm is close to perfect: plenty of parallelism and a minimum of communication. There is only one drawback: it runs slowly on a multicore processor, as shown in Figure 2.
Memory Wall Part 2 - Fig 2

Typically, the array size used with Gauss-Seidel is too large to fit in a multicore processor cache. Each iteration will force the entire array to be read from memory. Actually, for the red/black scheme, the array will have to be read twice per iteration, first during the red updates and then during the black updates. This will quickly saturate the DRAM bandwidth and limit the performance on a multicore processor.

Finding the Door in the Memory Wall

Instead of just maximizing parallelism, we could try to minimize DRAM bandwidth usage for a GS implementation. If we apply a blocking scheme to the original GS algorithm, we could keep an active subset of the array, called a block, in the cache and reuse these elements many times before the data are evicted from the cache. Because of the data dependence of the original GS algorithm (the red arrows in Figure 1a) we have to apply a sliding blocking technique, shown in Figure 3a. The active block inside the red frame shown includes three rows. Once the next iteration values for all the elements in the block have been updated, as shown in 3a, the block is slid down one row, as shown in Figure 3b, and the next iteration values for those elements are updated. This improves the reuse of element values while they reside in the cache. Using this scheme, each element of the array will advance three iterations per sweep, which means that the array is only read from DRAM every third iteration. This implies that only one sixth of the DRAM bandwidth is needed compared with the red/black algorithm.

If the number of rows in the active block increases, even less DRAM bandwidth will be needed. Figure 4 shows the relationship between bandwidth usage and block size, as shown by the ThreadSpotter tool for different cache sizes. A typical last-level multicore cache is in the 2-12 Mbyte range. Bandwidth demand can be reduced in this range by more than an order of magnitude using the blocking GS scheme instead of the red/black GS scheme.

Figure 3c shows a parallel version of the blocked GS. A drawback is that the threads will have to synchronize row-wise to make sure the thread to the left stays slightly ahead of the thread to the right. In sum, the blocked GS algorithm produces about an order of magnitude more thread communication than red/black GS. Also, its parallelism is much worse, in the order of N parallel threads (one per column) can help out simultaneously, compared with the N2/2 parallelism of the red/black algorithm. Still, it outperforms the red/black algorithm by a factor of three on a dual-socket quad-core system thanks to the much lower DRAM bandwidth need. Similar results have been observed when comparing 3D versions of the algorithms. Figure 5 compares the performance of the two algorithms when running on a two-socket quad-core system. The red/black saturates the bandwidth already at two active cores, while the sliding GS algorithm scales well even on a two-socket system without any special thread applied.

Summing Up

In Part 1 of this article series, we saw how a throughput workload created a superlinear slowdown on a multicore architecture due to increased cache pressure, and in this article — Part 2 of the series — we were forced to change “the ideal” highly parallel and low-communication algorithm in order for it to run well on a multicore processor. This once more drives home the point made by Sanjiv Shah a couple of weeks ago: only focusing on parallelism is not always the best way to get good performance on a multicore architecture. In Part 3, we will take a look at various techniques for identifying when optimizations are needed and compare a few simple optimization tricks.

About the Author

Erik Hagersten is chief technology officer at Acumem, a Swedish-based company that offers performance analysis tools for modern processors. He was the chief architect for high-end servers at Sun Microsystems (the former Thinking Machines development team) for six years before moving back to Sweden in 1999. Erik was a consultant to Sun until Acumem started in 2006. Since 2000 his research team at Uppsala University has developed the key technology behind Acumem.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

TACC Helps ROSIE Bioscience Gateway Expand its Impact

April 26, 2017

Biomolecule structure prediction has long been challenging not least because the relevant software and workflows often require high end HPC systems that many bioscience researchers lack easy access to. Read more…

By John Russell

Messina Update: The U.S. Path to Exascale in 16 Slides

April 26, 2017

Paul Messina, director of the U.S. Exascale Computing Project, provided a wide-ranging review of ECP’s evolving plans last week at the HPC User Forum. Read more…

By John Russell

IBM, Nvidia, Stone Ridge Claim Gas & Oil Simulation Record

April 25, 2017

IBM, Nvidia, and Stone Ridge Technology today reported setting the performance record for a “billion cell” oil and gas reservoir simulation. Read more…

By John Russell

ASC17 Makes Splash at Wuxi Supercomputing Center

April 24, 2017

A record-breaking twenty student teams plus scores of company representatives, media professionals, staff and student volunteers transformed a formerly empty hall inside the Wuxi Supercomputing Center into a bustling hub of HPC activity, kicking off day one of 2017 Asia Student Supercomputer Challenge (ASC17). Read more…

By Tiffany Trader

HPE Extreme Performance Solutions

Remote Visualization Optimizing Life Sciences Operations and Care Delivery

As patients continually demand a better quality of care and increasingly complex workloads challenge healthcare organizations to innovate, investing in the right technologies is key to ensuring growth and success. Read more…

Groq This: New AI Chips to Give GPUs a Run for Deep Learning Money

April 24, 2017

CPUs and GPUs, move over. Thanks to recent revelations surrounding Google’s new Tensor Processing Unit (TPU), the computing world appears to be on the cusp of a new generation of chips designed specifically for deep learning workloads. Read more…

By Alex Woodie

Musk’s Latest Startup Eyes Brain-Computer Links

April 21, 2017

Elon Musk, the auto and space entrepreneur and severe critic of artificial intelligence, is forming a new venture that reportedly will seek to develop an interface between the human brain and computers. Read more…

By George Leopold

MIT Mathematician Spins Up 220,000-Core Google Compute Cluster

April 21, 2017

On Thursday, Google announced that MIT math professor and computational number theorist Andrew V. Sutherland had set a record for the largest Google Compute Engine (GCE) job. Sutherland ran the massive mathematics workload on 220,000 GCE cores using preemptible virtual machine instances. Read more…

By Tiffany Trader

NERSC Cori Shows the World How Many-Cores for the Masses Works

April 21, 2017

As its mission, the high performance computing center for the U.S. Department of Energy Office of Science, NERSC (the National Energy Research Supercomputer Center), supports a broad spectrum of forefront scientific research across diverse areas that includes climate, material science, chemistry, fusion energy, high-energy physics and many others. Read more…

By Rob Farber

Messina Update: The U.S. Path to Exascale in 16 Slides

April 26, 2017

Paul Messina, director of the U.S. Exascale Computing Project, provided a wide-ranging review of ECP’s evolving plans last week at the HPC User Forum. Read more…

By John Russell

ASC17 Makes Splash at Wuxi Supercomputing Center

April 24, 2017

A record-breaking twenty student teams plus scores of company representatives, media professionals, staff and student volunteers transformed a formerly empty hall inside the Wuxi Supercomputing Center into a bustling hub of HPC activity, kicking off day one of 2017 Asia Student Supercomputer Challenge (ASC17). Read more…

By Tiffany Trader

Groq This: New AI Chips to Give GPUs a Run for Deep Learning Money

April 24, 2017

CPUs and GPUs, move over. Thanks to recent revelations surrounding Google’s new Tensor Processing Unit (TPU), the computing world appears to be on the cusp of a new generation of chips designed specifically for deep learning workloads. Read more…

By Alex Woodie

NERSC Cori Shows the World How Many-Cores for the Masses Works

April 21, 2017

As its mission, the high performance computing center for the U.S. Department of Energy Office of Science, NERSC (the National Energy Research Supercomputer Center), supports a broad spectrum of forefront scientific research across diverse areas that includes climate, material science, chemistry, fusion energy, high-energy physics and many others. Read more…

By Rob Farber

Hyperion (IDC) Paints a Bullish Picture of HPC Future

April 20, 2017

Hyperion Research – formerly IDC’s HPC group – yesterday painted a fascinating and complicated portrait of the HPC community’s health and prospects at the HPC User Forum held in Albuquerque, NM. HPC sales are up and growing ($22 billion, all HPC segments, 2016). Read more…

By John Russell

Knights Landing Processor with Omni-Path Makes Cloud Debut

April 18, 2017

HPC cloud specialist Rescale is partnering with Intel and HPC resource provider R Systems to offer first-ever cloud access to Xeon Phi "Knights Landing" processors. The infrastructure is based on the 68-core Intel Knights Landing processor with integrated Omni-Path fabric (the 7250F Xeon Phi). Read more…

By Tiffany Trader

CERN openlab Explores New CPU/FPGA Processing Solutions

April 14, 2017

Through a CERN openlab project known as the ‘High-Throughput Computing Collaboration,’ researchers are investigating the use of various Intel technologies in data filtering and data acquisition systems. Read more…

By Linda Barney

DOE Supercomputer Achieves Record 45-Qubit Quantum Simulation

April 13, 2017

In order to simulate larger and larger quantum systems and usher in an age of “quantum supremacy,” researchers are stretching the limits of today’s most advanced supercomputers. Read more…

By Tiffany Trader

Google Pulls Back the Covers on Its First Machine Learning Chip

April 6, 2017

This week Google released a report detailing the design and performance characteristics of the Tensor Processing Unit (TPU), its custom ASIC for the inference phase of neural networks (NN). Read more…

By Tiffany Trader

Quantum Bits: D-Wave and VW; Google Quantum Lab; IBM Expands Access

March 21, 2017

For a technology that’s usually characterized as far off and in a distant galaxy, quantum computing has been steadily picking up steam. Read more…

By John Russell

Trump Budget Targets NIH, DOE, and EPA; No Mention of NSF

March 16, 2017

President Trump’s proposed U.S. fiscal 2018 budget issued today sharply cuts science spending while bolstering military spending as he promised during the campaign. Read more…

By John Russell

HPC Compiler Company PathScale Seeks Life Raft

March 23, 2017

HPCwire has learned that HPC compiler company PathScale has fallen on difficult times and is asking the community for help or actively seeking a buyer for its assets. Read more…

By Tiffany Trader

Nvidia Responds to Google TPU Benchmarking

April 10, 2017

Nvidia highlights strengths of its newest GPU silicon in response to Google's report on the performance and energy advantages of its custom tensor processor. Read more…

By Tiffany Trader

CPU-based Visualization Positions for Exascale Supercomputing

March 16, 2017

In this contributed perspective piece, Intel’s Jim Jeffers makes the case that CPU-based visualization is now widely adopted and as such is no longer a contrarian view, but is rather an exascale requirement. Read more…

By Jim Jeffers, Principal Engineer and Engineering Leader, Intel

For IBM/OpenPOWER: Success in 2017 = (Volume) Sales

January 11, 2017

To a large degree IBM and the OpenPOWER Foundation have done what they said they would – assembling a substantial and growing ecosystem and bringing Power-based products to market, all in about three years. Read more…

By John Russell

TSUBAME3.0 Points to Future HPE Pascal-NVLink-OPA Server

February 17, 2017

Since our initial coverage of the TSUBAME3.0 supercomputer yesterday, more details have come to light on this innovative project. Of particular interest is a new board design for NVLink-equipped Pascal P100 GPUs that will create another entrant to the space currently occupied by Nvidia's DGX-1 system, IBM's "Minsky" platform and the Supermicro SuperServer (1028GQ-TXR). Read more…

By Tiffany Trader

Leading Solution Providers

Tokyo Tech’s TSUBAME3.0 Will Be First HPE-SGI Super

February 16, 2017

In a press event Friday afternoon local time in Japan, Tokyo Institute of Technology (Tokyo Tech) announced its plans for the TSUBAME3.0 supercomputer, which will be Japan’s “fastest AI supercomputer,” Read more…

By Tiffany Trader

Is Liquid Cooling Ready to Go Mainstream?

February 13, 2017

Lost in the frenzy of SC16 was a substantial rise in the number of vendors showing server oriented liquid cooling technologies. Three decades ago liquid cooling was pretty much the exclusive realm of the Cray-2 and IBM mainframe class products. That’s changing. We are now seeing an emergence of x86 class server products with exotic plumbing technology ranging from Direct-to-Chip to servers and storage completely immersed in a dielectric fluid. Read more…

By Steve Campbell

IBM Wants to be “Red Hat” of Deep Learning

January 26, 2017

IBM today announced the addition of TensorFlow and Chainer deep learning frameworks to its PowerAI suite of deep learning tools, which already includes popular offerings such as Caffe, Theano, and Torch. Read more…

By John Russell

BioTeam’s Berman Charts 2017 HPC Trends in Life Sciences

January 4, 2017

Twenty years ago high performance computing was nearly absent from life sciences. Today it’s used throughout life sciences and biomedical research. Genomics and the data deluge from modern lab instruments are the main drivers, but so is the longer-term desire to perform predictive simulation in support of Precision Medicine (PM). There’s even a specialized life sciences supercomputer, ‘Anton’ from D.E. Shaw Research, and the Pittsburgh Supercomputing Center is standing up its second Anton 2 and actively soliciting project proposals. There’s a lot going on. Read more…

By John Russell

Facebook Open Sources Caffe2; Nvidia, Intel Rush to Optimize

April 18, 2017

From its F8 developer conference in San Jose, Calif., today, Facebook announced Caffe2, a new open-source, cross-platform framework for deep learning. Caffe2 is the successor to Caffe, the deep learning framework developed by Berkeley AI Research and community contributors. Read more…

By Tiffany Trader

HPC Startup Advances Auto-Parallelization’s Promise

January 23, 2017

The shift from single core to multicore hardware has made finding parallelism in codes more important than ever, but that hasn’t made the task of parallel programming any easier. Read more…

By Tiffany Trader

HPC Technique Propels Deep Learning at Scale

February 21, 2017

Researchers from Baidu’s Silicon Valley AI Lab (SVAIL) have adapted a well-known HPC communication technique to boost the speed and scale of their neural network training and now they are sharing their implementation with the larger deep learning community. Read more…

By Tiffany Trader

IDG to Be Bought by Chinese Investors; IDC to Spin Out HPC Group

January 19, 2017

US-based publishing and investment firm International Data Group, Inc. (IDG) will be acquired by a pair of Chinese investors, China Oceanwide Holdings Group Co., Ltd. Read more…

By Tiffany Trader

  • arrow
  • Click Here for More Headlines
  • arrow
Share This