A Case for a More Effective, Power-Efficient Turbo Boosting

Abstract

Single-thread performance and throughput often pose different design constraints and require compromises. Mainstream CPUs today incorporate a non-trivial number of cores, even for mobile devices. For power and thermal considerations, by default, a single core does not operate at the maximum performance level. When operating conditions allow, however, commercial products often rely on turbo boosting, which temporarily increases the clock frequency to increase single-thread performance. However, increasing clock speed may result in a poor performance return for invested energy. In this article, we make a case for a more effective boosting strategy, which invests energy in activities with the best estimated return. In addition to running faster clocks, we can also use a look-ahead thread to overlap the penalties of cache misses and branch mispredicts. Overall, for similar power consumptions, the proposed adaptive turbo boosting strategy can achieve about twice the performance benefits while halving the energy overhead. CCS Concepts: • Computer systems organization → Multicore architectures; Heterogeneous (hybrid) systems; • Hardware → Chip-level power issues;

R3-DLA (Reduce, Reuse, Recycle): A More Efficient Approach to Decoupled Look-Ahead Architectures

“…On the other extreme of the spectrum, an idle core in a multicore system is used to execute a different copy of the original program on a separate thread context [5], [7], [34]- [36], [57], [58], [60], [70], [83], [84], [101], [112]. This copy is often a reduced version of the program (which we referred to as the skeleton) so that it can run faster to look ahead.…”

Bootstrapping

“…On the other extreme of the spectrum, an idle core in a multicore system is used to execute a dierent copy of the original program on a separate thread context [3,[7][8][9][39][40][41][42][43][44][45][46][47]. This copy is often a reduced version of the program (which we referred to as the skeleton) so that it can run faster to look ahead.…”

“…The segment between two neighboring instances of the same eigen-branch is, for our purposes, an iteration. 7 Depending on the eigen-branch, a single iteration may be too short to accurately measure its behavior. We thus use a period of an integer number of iterations longer than a certain length, say, 10,000 dynamic instructions.…”

“…Even in the occasional displacement of a state for an eigen-branch, a median loop lasts over 5 million instructions, 30 times longer than the time needed for testing. 7 Segments between two neighboring but dierent eigen-branches are ignored because they are statistically insignicant. (less than 0.001%) Alternatives: Note that there are other proposals to track program phase behavior based on xed lengths [64].…”

“…Second, assuming energy doubles is a gross exaggeration, as we will discuss in more detail later. Indeed, DLA is already competitive against some current practices that boost single-thread performance [7]. In particular, using a wide-issue superscalar is one such practice.…”