Compilers and More: What To Do With All Those Cores?

By Michael Wolfe

April 6, 2007

I just returned from two small conferences, CGO (Code Generation and Optimization, and PPoPP (Principles and Practice of Parallel Programming, One of the key themes in both conferences, perhaps the dominating theme, was that multicore chips are here, are mainstream, and we’d better figure out how to use them.

One of the PPoPP attendees, Prof. Rudolf Eigenmann (Purdue Univ.) issued an indictment, saying that we in the parallel programming research community should be ashamed of ourselves. Single-processor systems have run out of steam, something the parallel programming community has been predicting since I was a college student. Now is the time to step up and reap the benefits of all our past work. We’ve had 30 years to study this problem and come up with a solution, but what’s the end result? Surprise! We still have no well-accepted method to generate parallel applications.

We had a similar discussion at one of the workshops before CGO, where someone said that the programming problem would finally get solved because “now we’re motivated!” This implies either that the past thirty years of investigation was being done by the wrong people, or that perhaps now with more people looking at the problem, someone will randomly stumble on a good solution. Science at its best?

Dr. Andrew Chien (Intel), one of the PPoPP keynote speakers, took issue with Eigenmann’s criticism. Chien said that in fact we’ve had a great deal of success in parallel programming: just look at all the massively parallel systems and the applications that run on them. However, halfway through his talk was the slide “Wanted: Breakthrough Innovations in Parallel Programming.” I asked how he could claim past success, then state that breakthrough innovations are needed; it sounded like a typical manager: “good job, now get back to work.” He replied that in the past, parallel programming meant high performance. Now, parallel programming means spreadsheets, games, email, and applications on your laptop. It’s a different target environment, with a different class of programmer, and different expectations.

So parallel programming is hard. Hey, sequential programming is hard too. Adding parallelism just makes it harder. Two of the major problems are expressing the parallelism and synchronizing or communicating between the parallel threads or activities.

An approach to the synchronization problem that is gaining lots of traction is transactional memory, or TM. This takes a page from the successful database community, which has long had to deal with many processes simultaneously accessing and modifying a shared database. Rather than letting a process lock the database, the process executes a transaction; the transaction may involve adding, removing, or modifying relations in the database. While in the middle of a transaction, no other process can see the modifications; at the end, the process commits the results. The database then atomically exposes all the updates to the other processes. In particular, this lets multiple processes modify disjoint parts of the database in parallel without conflict. If two processes try to update the same data at the same time, the first commit will succeed, and the second will fail. The application then has to restart its transaction from the start, since it may have made some decisions based on values that are now stale.

Moving transactions to the parallel programming world means rather than updating shared data in a critical section, the model is to enter a transaction, perform the updates, then commit the changes. The implementation must buffer the modifications until the commit, then atomically commit all the modifications at once. As in a database, if some other parallel thread had made changes to the same shared data, the commit will fail, and the transaction must restart and retry. The expected behavior is that most transactions will succeed, so the retry overhead is quite low.

However, it’s still open as to how to implement the transaction buffering and commit. Modified caches could be used to buffer stores, unless you run out of associativity on the cache line. A separate transactional store queue could be designed, but the size of the queue would limit the size of the transaction.

There are other problems as well. Imagine two transactions, A and B, reading and modifying the same shared data. Suppose transaction B finishes and commits while A is still working. If transaction A then reads some data that was modified by B, A’s commit is likely to fail (in fact, may be guaranteed to fail), since some of the data it read was before B’s commit and some was read after. In fact, the inconsistency may cause A to generate a fault (suppose B allocated or freed a pointer), or loop infinitely; one speaker termed A a “zombie transaction,” the walking dead. It can be important to detect and kill zombies before they get to the commit state, to avoid spurious faults.

In managed software environments (think Java or C#), these problems can be handled in software, and transactions are likely to be successful there. However, it remains unclear how long it will be before transactions can migrate into HPC.

The architectural trends for multicore processing are still in flux as well. One idea is to build an array of small, low-power cores on the chip. Each core is slower on a single thread, but the array of cores could provide a great deal of job throughput. Andrew Chien pointed out that the single stream performance of a core grows roughly relative to the square root of the area, whereas the aggregate performance of a multicore chip grows linearly with the number of cores — a strong argument.

The Sun UltraSPARC T1 (Niagara) processor uses this strategy, with eight cores, each core swapping between four threads (much like the PPUs of the Control Data 6600, back in the late 1960s), giving very high throughput with significant power savings. This sacrifices single-thread performance, and Amdahl’s Law says this will affect parallel applications performance as well.

To solve that problem, some proposed designs use one or two aggressive, out-of-order, deeply pipelined, superscalar (large, power-hungry) cores surrounded by a sea of more tame, in-order, (smaller, power-frugal) cores. A parallel application would run mostly on the sea of slower cores, with the larger cores powered down. When entering a large sequential region on the critical path, the program would shift to the more powerful core.

This idea is not new, I first heard something similar 15 years ago or more, from Samuel Ho, a young PhD candidate at the University of Washington. At that time, he was proposing an Intel 486 surrounded by a bunch of 386s (which gives you the time frame), mostly for reduced cost. Today’s reduced power arguments are more compelling.

In some arenas, multicore has a well-established history. High end graphics processing units (GPUs) in personal computers and workstations have for years exposed a great deal of parallelism, and now look more like a collection of programmable processors with additional functional units, data types, and instructions geared for graphics problems. There are new efforts to expose GPU programming to the more general-purpose high performance market — the so-called GPGPU programming.

However, there are some important caveats. For instance, today’s GPUs implement floating point arithmetic, but only 32-bit precision, and not all the IEEE rounding modes. Some of the operations are not precise to the last ULP (Unit in the Last Place). That much precision isn’t needed in the graphics world, and this simplifies the GPU, making it smaller and faster.

The hardest nut to swallow is the graphics memory, which right now doesn’t implement ECC (error correcting code). With close to a gigabyte of memory, transient single-bit errors are quite possible, but again, in the graphics world, that will likely correspond to a slightly off-shaded pixel somewhere on the screen for one frame, so who cares? In your numerical simulation, that one bit could be a little more important, so perhaps this isn’t quite ready for life-critical applications.

My summary of all the hype for GPGPUs is that processors or coprocessors unconstrained by compatibility requirements, with the freedom to redesign to the latest technology, can deliver higher performance than general purpose CPUs. This is something we’ve known since the days of Floating Point Systems, with its attached FP processors. There are other coprocessors specifically designed for the high performance market, such as the Clearspeed board. GPUs are convenient for the budget-minded, because the development cost is paid for by the consumer games market.

So that’s what we have, and it’s up to you to figure out how to use it. High performance computing, as usual, is left with the crumbs off the table of the mass market.

Right now, we mostly think of multicore chips with 2 or 4 cores, even as Sun prepares the UltraSPARC T2 with 8 cores for release this fall. The sweet spot of high performance parallel computing has been in the range of 4 to 32 processors. Many applications do scale up beyond that, but hardware becomes much more complex to deliver scalable communication bandwidth, and software must be restructured to take advantage of all the parallelism without spending all its time waiting on remote data.

David Callahan (Microsoft Research) pointed out that exponential grows really fast. If we plan on doubling the number of cores on a chip every year or 18 months, it won’t be long before we have hundreds or thousands of cores on a chip. Any software solution aimed at 16 or 32 cores will quickly become irrelevant. We’d better be looking at productive ways to use massively parallel systems, since these may well find their way into our workstations, and yes, even our laptops, before the end of the next decade.

At one CGO panel, the moderator asked the audience whether new languages were needed for multicore. The response was somewhat tepid. It’s not clear that multicore is the reason for developing new languages. There are several research projects looking at designing or modifying languages specifically for developing scalable parallel applications.

Several of these project involve the so-called PGAS (Partitioned Global Address Space) languages. The main idea is the data is explicitly and visibly partitioned across the parallel machine, but any thread running on any node can fetch or store remote data directly. This promotes MPI-style programming, with most of the computation on local data, but exposes the communication as a language primitive instead of using opaque library calls. They try to get the advantages of HPF (High Performance Fortran), such as global addressability, without relying on unproven compiler technology.

DARPA, as part of its High Productivity Computing Systems (HPCS) project, is sponsoring the development of three new languages; see the HPCwire Q&A with Rusty Lusk for a summary. Recently I took a tutorial of IBM’s HPCS language, X10. It includes some of the PGAS ideas for parallelism, and allows for several ways to express parallelism at multiple levels beyond just data parallelism. It will be quite interesting to watch the evolution of these languages, but if this tutorial is indicative of their maturity, we have years of development before they are ready for real users.

So what’s the answer to the question in this article’s title? What DO we do with all those cores? The gold standard which we would like to achieve is to make them invisible. Upgrading to a multicore processor should be as simple and as effective as upgrading to a faster processor was in years past.

Previously, improvements to microarchitectural elements, such as branch prediction, out-of-order execution control units and pipelined functional units, were, for the most part, invisible from software. Instruction-set architectural changes, such as the SSE and SSE2 registers and instructions, were visible to the compiler, but largely hidden from the programmer.

Some architectural enhancements are significant enough that changing the program to take advantage of them is worth the effort. The packed SSE instructions on x86 processors and the vector instructions of classical supercomputers fall into this category. If we can make multicore processors no harder to use than this, we will have succeeded. The biggest barrier is that the architecture and the OS today present the multicore processor as indistinguishable from a multiprocessor. We need hardware and software mechanisms to allow compilers and developers to take advantage of these multicore chips as powerful processors. That is the next short-term challenge and will be the subject of a future column.


Michael Wolfe has developed compilers for over 30 years in both academia and industry, and is now a senior compiler engineer at The Portland Group, Inc. (, a wholly-owned subsidiary of STMicroelectronics, Inc. The opinions stated here are those of the author, and do not represent opinions of The Portland Group, Inc. or STMicroelectronics, Inc.

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!

China’s Expanding Effort to Win in Microchips

July 27, 2017

The global battle for preeminence, or at least national independence, in semiconductor technology and manufacturing continues to heat up with Europe, China, Japan, and the U.S. all vying for sway. A fascinating article ( Read more…

By John Russell

Hyperion: Storage to Lead HPC Growth in 2016-2021

July 27, 2017

Global HPC external storage revenues will grow 7.8% over the 2016-2021 timeframe according to an updated forecast released by Hyperion Research this week. HPC server sales, by comparison, will grow a modest 5.8% to $14.8 Read more…

By John Russell

Exascale FY18 Budget – The Senate Provides Their Input

July 27, 2017

In the federal budgeting world, “regular order” is a meaningful term that is fondly remembered by members of both the Congress and the Executive Branch. Regular order is the established process whereby an Administrat Read more…

By Alex R. Larzelere

HPE Extreme Performance Solutions

HPE Servers Deliver High Performance Remote Visualization

Whether generating seismic simulations, locating new productive oil reservoirs, or constructing complex models of the earth’s subsurface, energy, oil, and gas (EO&G) is a highly data-driven industry. Read more…

India Plots Three-Phase Indigenous Supercomputing Strategy

July 26, 2017

Additional details on India's plans to stand up an indigenous supercomputer came to light earlier this week. As reported in the Indian press, the Rs 4,500-crore (~$675 million) supercomputing project, approved by the Ind Read more…

By Tiffany Trader

Exascale FY18 Budget – The Senate Provides Their Input

July 27, 2017

In the federal budgeting world, “regular order” is a meaningful term that is fondly remembered by members of both the Congress and the Executive Branch. Reg Read more…

By Alex R. Larzelere

India Plots Three-Phase Indigenous Supercomputing Strategy

July 26, 2017

Additional details on India's plans to stand up an indigenous supercomputer came to light earlier this week. As reported in the Indian press, the Rs 4,500-crore Read more…

By Tiffany Trader

Tuning InfiniBand Interconnects Using Congestion Control

July 26, 2017

InfiniBand is among the most common and well-known cluster interconnect technologies. However, the complexities of an InfiniBand (IB) network can frustrate the Read more…

By Adam Dorsey

NSF Project Sets Up First Machine Learning Cyberinfrastructure – CHASE-CI

July 25, 2017

Earlier this month, the National Science Foundation issued a $1 million grant to Larry Smarr, director of Calit2, and a group of his colleagues to create a comm Read more…

By John Russell

Graphcore Readies Launch of 16nm Colossus-IPU Chip

July 20, 2017

A second $30 million funding round for U.K. AI chip developer Graphcore sets up the company to go to market with its “intelligent processing unit” (IPU) in Read more…

By Tiffany Trader

Fujitsu Continues HPC, AI Push

July 19, 2017

Summer is well under way, but the so-called summertime slowdown, linked with hot temperatures and longer vacations, does not seem to have impacted Fujitsu's out Read more…

By Tiffany Trader

Researchers Use DNA to Store and Retrieve Digital Movie

July 18, 2017

From abacus to pencil and paper to semiconductor chips, the technology of computing has always been an ever-changing target. The human brain is probably the com Read more…

By John Russell

The Exascale FY18 Budget – The Next Step

July 17, 2017

On July 12, 2017, the U.S. federal budget for its Exascale Computing Initiative (ECI) took its next step forward. On that day, the full Appropriations Committee Read more…

By Alex R. Larzelere

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 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

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. Just how close real-wo 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 a Read more…

By Tiffany Trader

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 cam Read more…

By John Russell

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

Nvidia’s Mammoth Volta GPU Aims High for AI, HPC

May 10, 2017

At Nvidia's GPU Technology Conference (GTC17) in San Jose, Calif., this morning, CEO Jensen Huang announced the company's much-anticipated Volta architecture a Read more…

By Tiffany Trader

How ‘Knights Mill’ Gets Its Deep Learning Flops

June 22, 2017

Intel, the subject of much speculation regarding the delayed, rewritten or potentially canceled “Aurora” contract (the Argonne Lab part of the CORAL “ Read more…

By Tiffany Trader

Leading Solution Providers

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

Reinders: “AVX-512 May Be a Hidden Gem” in Intel Xeon Scalable Processors

June 29, 2017

Imagine if we could use vector processing on something other than just floating point problems.  Today, GPUs and CPUs work tirelessly to accelerate algorithms Read more…

By James Reinders

Russian Researchers Claim First Quantum-Safe Blockchain

May 25, 2017

The Russian Quantum Center today announced it has overcome the threat of quantum cryptography by creating the first quantum-safe blockchain, securing cryptocurrencies like Bitcoin, along with classified government communications and other sensitive digital transfers. Read more…

By Doug Black

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

Google Debuts TPU v2 and will Add to Google Cloud

May 25, 2017

Not long after stirring attention in the deep learning/AI community by revealing the details of its Tensor Processing Unit (TPU), Google last week announced the Read more…

By John Russell

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 Read more…

By Alex Woodie

Six Exascale PathForward Vendors Selected; DoE Providing $258M

June 15, 2017

The much-anticipated PathForward awards for hardware R&D in support of the Exascale Computing Project were announced today with six vendors selected – AMD Read more…

By John Russell

Top500 Results: Latest List Trends and What’s in Store

June 19, 2017

Greetings from Frankfurt and the 2017 International Supercomputing Conference where the latest Top500 list has just been revealed. Although there were no major Read more…

By Tiffany Trader

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