Will it blend? Merging heterogeneous cores

Abstract

Heterogeneous processors allow different performance/power operation points by pairing high performance Outof-Order (OoO) cores with small energy efficient In-Order cores. Timely switching between the cores types allows the system to tailor to energy budget and performance constraints. Each core migration incurs a penalty though, which is caused by the necessary transfer of state between cores. Heterogeneous systems are kept from further increasing performance through more frequent migrations, as the benefits are suppressed by these overheads. Sharing some of the components between cores can reduce the transfer overheads and potentially overcome this problem. However, multiplexing core components can have also an adverse effect on performance when done without taking into consideration the extra latency. For this reason, it is critical to strike a balance between component sharing and complexity.
In our work, we show that sharing the translation mechanism and the level 2 caches capture most of the potential benefits for in-order cores, while the branch predictor state is most critical for Out-of-Order cores. Finally, even under a best case scenario, on average 10% of applications could transfer to an OoO core to gain extra performance, while 6% of execution could migrate to In-Order cores saving significant energy, without incurring any penalty to the system. Under a more relaxed trade-off policy, benefits increase significantly as, on average, for 66% of the time executed a migration to the opposite core to improve the energy delay product.

More information

Identifiers

ORCID for Ilias Vougioukas: orcid.org/0000-0003-1444-4326

ORCID for Geoff Merrett: orcid.org/0000-0003-4980-3894

Catalogue record

Export record