Distributed in-network task scheduling for datacenters

Scheduling latency-sensitive applications in large-scale datacenters is challenging. Current approaches use application-layer schedulers, which impose high overheads and result in long latencies. We present Saqr, the first in-network, datacenter-wide scheduler that supports short tasks with execution times in the order of tens of microseconds. Saqr introduces new network-level constructs and a distributed scheduling policy to enable network switches to efficiently schedule tasks within the network at line rate and with minimal latency. We implemented Saqr in a testbed with high-speed programmable switches and compared its performance against the state-of-the-art in-network scheduler (Racksched). Our results show that Saqr can reduce the tail response time by up to 85% and the processing load on switches by up to 2.5X compared to Racksched. In addition, we compared Saqr versus Racksched using large-scale simulations with diverse and dynamic workloads and our results show that Saqr substantially outperforms Racksched across all performance metrics.