Abstract

Extensive data analysis has become the enabler for diagnostics and decision making in many modern systems. These analyses have both competitive as well as social benefits. To cope with the deluge in data that is growing faster than Moore’s law, computation frameworks have resorted to massive parallelization of analytics jobs into many fine-grained tasks. These frameworks promised to provide efficient and fault-tolerant execution of these tasks. However, meeting this promise in clusters spanning hundreds of thousands of machines is challenging and a key departure from earlier work on parallel computing. A simple but key aspect of parallel jobs is the all-or-nothing property: unless all tasks of a job are provided equal improvement, there is no speedup in the completion of the job. This talk will demonstrate how the all-or-nothing property impacts replacement algorithms in distributed caches for parallel jobs. Our coordinated caching system, PAC Man, makes global caching decisions and employs a provably optimal cache replacement algorithm. A highlight of our evaluation using workloads from Facebook and Bing datacenters is that PAC Man’s replacement algorithm outperforms even Belady’s MIN (that uses an oracle) in speeding up jobs. Along the way, I will also describe how we broke the myth of disk-locality’s importance in datacenter computing and solutions to mitigate straggler tasks.