Resumen de Performance and power optimizations in chip multiprocessors for throughput-aware computation
Augusto Javier Vega
The so-called "power (or power density) wall" has caused core frequency (and single-thread performance) to slow down, giving rise to the era of multi-core/multi-thread processors. For example, the IBM POWER4 processor, released in 2001, incorporated two single-thread cores into the same chip. In 2010, IBM released the POWER7 processor with eight 4-thread cores in the same chip, for a total capacity of 32 execution contexts. The ever increasing number of cores and threads gives rise to new opportunities and challenges for software and hardware architects. At software level, applications can benefit from the abundant number of execution contexts to boost throughput. But this challenges programmers to create highly-parallel applications and operating systems capable of scheduling them correctly. At hardware level, the increasing core and thread count puts pressure on the memory interface, because memory bandwidth grows at a slower pace ---phenomenon known as the "bandwidth (or memory) wall". In addition to memory bandwidth issues, chip power consumption rises due to manufacturers' difficulty to lower operating voltages sufficiently every processor generation. This thesis presents innovations to improve bandwidth and power consumption in chip multiprocessors (CMPs) for throughput-aware computation: a bandwidth-optimized last-level cache (LLC), a bandwidth-optimized vector register file, and a power/performance-aware thread placement heuristic. In contrast to state-of-the-art LLC designs, our organization avoids data replication and, hence, does not require keeping data coherent. Instead, the address space is statically distributed all over the LLC (in a fine-grained interleaving fashion). The absence of data replication increases the cache effective capacity, which results in better hit rates and higher bandwidth compared to a coherent LLC. We use double buffering to hide the extra access latency due to the lack of data replication. The proposed vector register file is composed of thousands of registers and organized as an aggregation of banks. We leverage such organization to attach small special-function "local computation elements" (LCEs) to each bank. This approach ---referred to as the "processor-in-regfile" (PIR) strategy--- overcomes the limited number of register file ports. Because each LCE is a SIMD computation element and all of them can proceed concurrently, the PIR strategy constitutes a highly-parallel super-wide-SIMD device (ideal for throughput-aware computation). Finally, we present a heuristic to reduce chip power consumption by dynamically placing software (application) threads across hardware (physical) threads. The heuristic gathers chip-level power and performance information at runtime to infer characteristics of the applications being executed. For example, if an application's threads share data, the heuristic may decide to place them in fewer cores to favor inter-thread data sharing and communication. In such case, the number of active cores decreases, which is a good opportunity to switch off the unused cores to save power. It is increasingly harder to find bulletproof (micro-)architectural solutions for the bandwidth and power scalability limitations in CMPs. Consequently, we think that architects should attack those problems from different flanks simultaneously, with complementary innovations. This thesis contributes with a battery of solutions to alleviate those problems in the context of throughput-aware computation: 1) proposing a bandwidth-optimized LLC; 2) proposing a bandwidth-optimized register file organization; and 3) proposing a simple technique to improve power-performance efficiency.
