دانلود رایگان مقاله ترافیک برنامه ریزی برای حداکثر طول عمر سوئیچ در مرکز داده های نوری

Abstract

Optical switching is a promising technology to scale the performance of data centers. We address two scenarios: hybrid data centers, in which an optical circuit switching network operates in parallel to an electronic packet switching network that interconnect the servers, and full optical data centers, in which all the traffic is switched in the optical domain. The traffic on the optical network is scheduled proactively by configuring the optical fabric according to a precomputed sequence of configurations. We consider that fast optical circuit switching is implemented through micro-mechanical devices (like MEMS) at each crosspoint, whose lifetime is limited by a maximum number of configuration cycles, due to the experienced mechanical fatigue. We consider the traffic scheduling problem on the optical network and address specifically the problem of minimizing the number of reconfigurations at each crosspoint, in order to maximize the MEMS lifetime, while maximizing the throughput. We propose a family of scheduling algorithms and discuss the achievable tradeoff between lifetime, throughput and computational complexity, throughout a set of extensive simulations and theoretical results.

7. Conclusions

We have considered an optical fabric to interconnects ToR switches in a data center, allowing fast reconfigurable circuit switching among the racks. We have assumed that MEMS are adopted for full optical switching, and this technology is affected by limited lifetime due to the mechanical fatigue. We addressed the problem of scheduling the traffic across racks to maximize the performance in terms of throughput and to maximize the lifetime of the switching fabric. The main idea is to minimize the number of single variations between consecutive switching configurations. We have proposed a family of fatigue-aware frame scheduling algorithms that offer different tradeoff between performance (throughput and fatigue) and computational complexity. We have investigated the achieved tradeoffs using both analytical and simulative approaches, under both synthetic and realistic traffic matrices observed in operational Facebook and Microsoft data centers. As result, we have shown that, in a majority of the scenarios, the specific algorithm denoted as GExa-NS, is outperforming all the other approaches based on state-of-art algorithms. Notably, GExa-NS can substitute the standard decomposition algorithms for the traffic scheduling in optical fabrics, without incurring in any extra computational cost, but providing an important improvement in the lifetime of the MEMS-based switches.