A Sleep-based Communication Mechanism to Save Processor Utilization in Distributed Streaming Systems

Energy-efficiency of applications deployed in data-centers is becoming increasingly important. Techniques that reduce CPU utilization for specific workloads can help improve energy consumption. An application domain that has been studied in the past extensively and is lately gaining importance in data-centers is distributed stream processing. In this work we examine an existing stream processing system, Borealis [6], and we identify significant sources of overhead in the communication stack. Specifically, we examine the inter-node communication path in a distributed setup and the overheads associated as streams flow from node to node. We find that the send and receive tasks in Borealis take up significant CPU resources. We redesign the send and receive paths of Borealis by replacing TCP with a user-level protocol based on Myrinet MX. We then evaluate the CPU utilization and network throughput on a 10 Gbits/s network using both polling and interrupts for communicating data. Finally, we propose a sleep mechanism that avoids the CPU overheads associated with both interrupts and polling. We use a real setup consisting of four eight-core nodes equipped with 10 Gbits/s Ethernet and native Myrinet MX communication subsystems to examine the impact of our approach. Our results show that our approach saves CPU utilization for a range of workload conditions and is able to achieve better throughput compared to TCP with lower CPU utilization (upto 40%).