CONSTRUCTING A FLEXIBLE INTERNET-SCALE TIME-SHARING SYSTEM USING DETERMINISTIC CHECKPOINTING

摘要

Distributed systems and clustering have both grown into mainstream phenomena, now used throughout academia and industry. Despite advances in a wide variety of interconnects, including home Internet connections at speeds and latencies only available in a data center several years ago, the scheduling and use of these systems is still firmly rooted in the era of batch processing. Batch processing provides a basic scheduler and simple programming model for many types of computation, but lacks the flexibility and efficient resource utilization of even the most rudimentary of time-sharing systems. With a large number of use cases now constrained by these limitations, time-sharing concepts must once again come to the rescue. Based on experience from the distributed.net, Folding@home, and Storage@home systems, it can be demonstrated that the Internet has advanced and can now meet the higher requirements for time-sharing. This paper will explore those requirements and the potential benefits of moving beyond batch processing and into time-sharing for Internet-scale computations, and lay out a method for deterministic checkpointing required to implement such a system.