Google is the modern data poster-child for parallel computing. It's famous for splintering enormous calculations into tiny pieces that can then be processed across an epic network of machines. But when it comes to spreading workloads across multi-core processors, the company has called for a certain amount of restraint.
With a paper (PDF) soon to be published in IEEE Micro, the IEEE magazine of chip and silicon design, Google Senior Vice President of Operations Urs Hölzle – one of the brains overseeing the web giant's famous back-end – warns against the use of multi-core processors that take parallelization too far. Chips that spread workloads across more energy-efficient but slower cores, he says, may not be preferable to chips with faster but power-hungry cores.
Hölzle sees this as the battle of the "wimpy" cores and the "brawny" cores.
"Slower but energy efficient 'wimpy' cores only win for general workloads if their single-core speed is reasonably close to that of mid-range 'brawny' cores," he says. The problem, he explains, is that wimpy cores run into Amdahl's law (PDF). In essence, Amdahl's law says that when you parallelize only part of a system, there is a limit to performance improvement.
"So why doesn’t everyone want wimpy-core systems?" Hölzle writes. "Because in many corners of the real world, they’re prohibited by law — Amdahl’s law. Even though many Internet services benefit from seemingly unbounded request- and data-level parallelism, such systems aren’t above the law. As the number of parallel threads increases, reducing serialization and communication overheads can become increasingly difficult. In a limit case, the amount of inherently serial work performed on behalf of a user request by slow single-threaded cores will dominate overall execution time."
When considering "wimpy" cores, he continues, you can't forget the cost of software development. "Wimpy-core systems can require applications to be explicitly parallelized or otherwise optimized for acceptable performance. For example, suppose a Web service runs with a latency of one second per user request, half of it caused by serial CPU time. If we switch to wimpy-core servers, whose single-threaded performance is three times slower, the response time doubles to two seconds and developers might have to spend a substantial amount of effort to optimize the code to get back to the one- second latency."
The other problem, he says, is that the more you parallelize, the more you increase response time. This is why Google's distributed number crunching platform, MapReduce, isn't suited to real-time calculations. "Often all parallel tasks must finish before a request is completed, and thus the overall response time becomes the maximum response time of any subtask, and more subtasks will push further into the long tail of subtask response times."
The use of wimpy servers can raise non-CPU hardware costs, he continues, and lower utilization. "Consider the task of allocating a set of applications across a pool of servers as a bin-packing problem — each of the servers is a bin, and we try to fit as many applications as possible into each bin. Clearly that task is harder when the bins are small, because many applications might not completely fill a server and yet use too much of its CPU or RAM to allow a second application to coexist on the same server."
Most surprisingly, Hölzle says that extreme parallelization can be less efficient when used on a, well, global scale. "To avoid expensive global communication and global lock contention, local tasks can use heuristics that are based on their local progress only, and such heuristics are naturally more conservative. As a result, local subtasks might execute for longer than they would have if better hints about global progress were available. Naturally, when these computations are partitioned into smaller pieces, this overhead tends to increase."
All this leads the Google man to conclude that spreading calculations across a larger collection of wimpy cores doesn't always make sense. "Although we’re enthusiastic users of multicore systems, and believe that throughput-oriented designs generally beat peak-performance-oriented designs, smaller isn’t always better," he says. "Once a chip’s single-core performance lags by more than a factor to two or so behind the higher end of current-generation commodity processors, making a business case for switching to the wimpy system becomes increasingly difficult because application programmers will see it as a significant performance regression: their single-threaded request handlers are no longer fast enough to meet latency targets.
"So go forth and multiply your cores, but do it in moderation, or the sea of wimpy cores will stick to your programmers’ boots like clay." ®