Section: New Results

Processor Architecture

Participants : Arthur Blanleuil, Niloofar Charmchi, Caroline Collange, Kleovoulos Kalaitzidis, Pierre Michaud, Anis Peysieux, Daniel Rodrigues Carvalho, André Seznec.

Value prediction

Participants : Kleovoulos Kalaitzidis, André Seznec.

Modern context-based value predictors tightly associate recurring values with instructions and contexts by building confidence upon them [9]. However, when execution monotony exists in the form of intervals, the potential prediction coverage is limited, since prediction confidence is reset at the beginning of each new interval. In [25], we address this challenge by introducing the notion of Equality Prediction (EP), which represents the binary facet of value prediction. Following a two fold decision scheme (similar to branch prediction), EP makes use of control-flow history to determine equality between the last committed result read at fetch time, and the result of the fetched occurrence. When equality is predicted with high confidence, the read value is used. Our experiments show that this technique obtains the same level of performance as previously proposed state-of-the-art context-based predictors. However, by virtue of better exploiting patterns of interval equality, our design complements the established way that value prediction is performed, and when combined with contemporary prediction models, improves the delivered speedup by 19 % on average.

Compressed caches

Participants : Daniel Rodrigues Carvalho, Niloofar Charmchi, Caroline Collange, André Seznec.

The speed gap between CPU and memory is impairing performance. Cache compression and hardware prefetching are two techniques that could confront this bottleneck by decreasing last level cache misses. However, compression and prefetching have positive interactions, as prefetching benefits from higher cache capacity and compression increases the effective cache size. We propose Compressed cache Layout Aware Prefetching (CLAP) to leverage the recently proposed sector-based compressed cache layouts such as SCC or YACC to create a synergy between compressed cache and prefetching. The idea of this approach is to prefetch contiguous blocks that can be compressed and co-allocated together with the requested block on a miss access [33]. Prefetched blocks that share storage with existing blocks do not need to evict a valid existing entry; therefore, CLAP avoids cache pollution. In order to decide the co-allocatable blocks to prefetch, we propose a compression predictor. Based on our experimental evaluations, CLAP reduces the number of cache misses by 12 % and improves performance by 4 % on average, comparing to a compressed cache [23].

Deep microarchitecture

Participants : Anis Peysieux, André Seznec.

The design of an efficient out-of-order execution core is particularly challenging. When the issue-width increases, the cost of the extra logic required by out-of-core execution increases dramatically. The silicon area occupied by this OoO core tends to grow quasi-quadratically with the issue-width (e.g. issue logic, register file and result bypass). At the same time, the power requirement and the energy consumption of the out-of–order core grow super-linearly with issue width. On wide-issue out-of-order execution cores, issue logic response time, register file access time, as well as result bypass delays represent potential critical paths that might impair cycle time or might necessitate further deepening of the execution pipeline. The objective of the PhD thesis of Anis Peysieux will be to reduce the number of instructions that enter the OoO core, and therefore to master the hardware complexity while still achieving the performance promises of a very wide issue processor.

Dynamic thermal management

Participant : Pierre Michaud.

As power dissipation and circuit temperature constrain their performance, modern processors feature turbo control mechanisms to adjust the voltage and clock frequency dynamically so that circuit temperature stays below a certain limit. In particular, turbo control exploits the fact that, after a long period of low processor activity, the thermal capacity of the chip, its package and the heatsink can absorb heat at a relatively fast rate during a certain time, before the temperature limit constrains that rate. Hence power dissipation can be temporarily boosted above the average sustainable value. The turbo control must monitor circuit temperature continuously to maximize the clock frequency. Temperature can be monitored by reading the integrated thermal sensors. However, making the clock frequency depend on thermal sensor readings implies that processor performance depends on ambient temperature. Yet this form of performance non-determinism is a problem for certain processor makers. A possible solution is to determine the clock frequency not from the true temperature but from a thermal model based on the nominal ambient temperature. Such model should be as accurate as possible in order to prevent sensor-based protection from triggering but sporadically, without hurting performance by overestimating temperature too much. The model should also be simple enough to provide calculated temperature in real time. We propose a thermal model possessing these qualities, and a new turbo control algorithm based on that model [37].

Thread convergence prediction for general-purpose SIMT architectures

Participants : Arthur Blanleuil, Caroline Collange.

GPUs group threads of SPMD programs in warps and synchronize them to execute the same instruction at the same time. This execution model, referred to as Single-Instruction, Multiple-Thread (SIMT), enables the use of energy-efficient SIMD execution units by factoring out control logic such as instruction fetch and decode pipeline stages for a whole warp. SIMT execution is the key enabler for the energy efficiency of GPUs. We seek to generalize the SIMT execution model to general-purpose superscalar cores.

As threads within a warp may follow different directions through conditional branches in the program, the warp must follow each taken path in turn, while disabling individual threads that do not participate. Following divergence, current GPU architectures attempt to restore convergence at the earliest program point following static annotations in the binary. However, this policy has been shown to be suboptimal in many cases, in which later convergence improves performance. In fact, optimal convergence points depend on dynamic program behavior, so static decisions are unable to capture them.

The goal of the thesis of Arthur Blanleuil is to design predictors that enable the microarchitecture to infer dynamic code behavior and place convergence points appropriately. Convergence predictors have analogies with branch predictors and control independence predictors studied in superscalar processor architecture, but they present one additional challenge: the thread runaway problem. Although a branch misprediction will be identified and repaired locally, a wrong thread scheduling decision may go unnoticed and delay convergence by thousands of instructions. To address the thread runaway problem, we plan to explore promise-based speculation and recovery strategies. When no information is available, we follow the traditional conservative earliest-convergence scheduling policy. Once the predictor has enough information to make a more aggressive prediction, it generates assumptions about the prediction. The microarchitecture then keeps checking dynamically whether the assumptions actually hold true in the near future. If assumptions turn out to be wrong, the prediction will be reconsidered by changing back priorities to conservative. Such promise-based speculation policies can address the thread runaway problem by fixing a bound on the worst-case performance degradation of an aggressive scheduling policy against the conservative baseline.

Accurate thread convergence policies will enable dynamic vectorization to adapt to application characteristics dynamically. They will both improve performance and simplify programming of many-core architectures by alleviating the need for advanced code tuning by expert programmers.

Exploring the design space of GPU architectures

Participants : Alexandre Kouyoumdjian, Caroline Collange.

We study tradeoffs in the internal organization of GPUs in the context of general-purpose parallel processing [35]. In particular, we analyze the performance impact of having a few wide streaming multiprocessors compared to many narrow ones. Although we find narrow configurations usually give higher performance for an equal number of execution units, they require more hardware resources and energy. On the other hand, our evaluation show that the optimal streaming multiprocessor width varies across applications. This study motivates adaptive GPU architectures that would support configurable internal organization.