Figure 5. Heterogeneous server configuration with 25BCE large cores and 1BCE small cores.

Arrival rate: λ

short request long request

Figure 6. Server configurations with 10BCE cores when dedicating (a) 10 resource units and (b) 70 resource units toward caching.

( C =

10) Cac he

(a)

Arrival rate: λ

Arrival rate: λ

Service time: Ts = 1/(µ√ 10)

(b)

( C =

70) Cach e

and fourth root of r. Beyond that point, optimal design favors smaller cores.

Core Heterogeneity

The previous section explored the trade-offs between powerful, brawny cores and power-efficient, wimpy cores. Neither type of core provides high efficiency across a wide range of QoS targets, raising several obvious questions, including: Should an architect combine multiple core types in the same system, as is already the norm in multi-core chips for mobile systems? How should architects determine the size of these cores? And at what ratio should they use them? Determining the right mix of large-versus-little cores, as well as devising schedulers that take advantage of heterogeneous cores, especially in the presence of heterogeneous load, has been a notably active topic of research in computer architecture in recent years. 5, 9, 15 Figure 4c shows the QPS under various QoS targets for a set of heterogeneous designs. In all cases, the system has two core configurations: small cores with U = 1, benefiting applications with relaxed QoS, and big cores with U = 25, benefiting applications with strict QoS. The system also receives two exponentially distributed input request streams, one with short and the other with long mean-service-time requests, and design a simple heterogeneity-aware scheduler that routes long requests to big cores and short requests to small cores. Requests are admitted to a single queue, as in Figure 5, and the ratio of long-to-short requests is, for now, 1: 1. Figure 5 starts with all big cores at the leftmost point of the x-axis, explores the heterogeneous space, and ends with all small cores at the rightmost point.

Finding 4. Figure 4c captures a surprising trend. For strict QoS targets, like 1 · Ts, homogeneous systems with all big cores achieve optimal performance. In contrast, for very relaxed QoS targets, like 100Ts, using all small cores achieves the best performance. However, for QoS targets in the middle (such as 10Ts), heterogeneous systems, coupled with heterogeneity-aware schedulers, outperform their homogeneous counterparts. This result is especially true when the ratio of big to small cores matches the ratio of long-to-short requests. Varying the request ratio affects