JENERGY: A Fault Tolerant Stateless Architecture for High Performance Computing

Abstract

Large scale HPC (high performance computing) applications require thousands of nodes for computing parallel scientific applications. At this scale, hardware and software failures, network congestion or disconnections are frequent faults experienced by compute nodes. This introduces high levels of volatility which reduces the Mean Time between Failures (MTBF) of the whole system down to hours or minutes. To deal with this kind of failure rates, traditional point-topoint transmission semantics can be ill-fitted and cumbersome to re-engineer to support distributed partial failures. In this paper, we propose an application dependent network design that focuses on the sustainability of High Performance Computing (HPC) applications using packet-switching-inspired statistical multiplexing of semantic data tuples and decoupled computations. We report the design and implementation of a distributed tuple space using Cassandra and Zookeeper for tunable spatial and temporal redundancies without negative impact on application performance. We detail the various failure scenarios that can be handled seamlessly by our system and provide a description of the advantages of Stateless Parallel Processing for HPC applications. We report the preliminary results on performance, reliability and overall application scalability. We found that our system can provide high levels of sustained performance, while providing a reliable computing architecture that can withstand a range of failure types without manual checkpoint-restart, in a portable and non-intrusive manner.

Measure and compare the convergence time of network routing protocols

“…The MTBF (Mean Time Between Failures) value for devices of critical infrastructure are getting longer and the reliability of a network devices (availability) reaches 99,999%. However this level of reliability, in some cases, is still not sufficient enough [1]. The formula of improving network availability by improving the device production process and improving the reliability parameters of network nodes seems to be reaching its limit.…”