Abstract

The continuous growth of data centers is making power efficient computation a rising concern. The ubiquity of large scale-out workloads on the other hand provides enough scale for specialization to become a viable option. In this talk I will present our work on achieving very high power efficiency without sacrificing performance for a common data center application by the means of specialization.

Key-value stores such as memcached are commonly used to reduce the demand on relational databases, which are much harder to scale. However, since these key-value stores are more network heavy than computationally intensive, they do not utilize modern server hardware efficiently. To overcome this inefficiency the use of smaller ARM -based machines in large numbers has been proposed, but our goal is to dramatically reduce the power consumption of key-value stores and similar applications without compromising performance. Therefor we took specialization a step further: We present a microserver that implements the target application (memcached) directly in hardware, fusing network handling and application-specific processing into a dataflow architecture on an FPGA . As a result, power consumption is reduced ten-fold when compared to regular servers. Moreover, the performance and response time of software-based solutions is matched and even surpassed in many cases.