Çok İşlemcili Ağlarda Paralel Ayrık Olay Benzetimlerinin Dağıtık İşletim Sistemi Donanımı için Anahtar-Tabanlı Yaklaşım

Abstract

Time advancement and publish/subscribe message distribution with minimum latency has the utmost importance for the performance of parallel discrete event simulations. Parallel discrete event simulations performance improvement is aimed at the designed network accelerator. High Level Architecture (HLA) is used as a parallel discrete event simulation infrastructure. Time management (TM) and data distribution management (DDM) services are provided by run time infrastructure (RTI). The TM and DDM services’ latency is improved to increase the simulation performance. TM service’s greatest available logic time (GALT) and DDM service’s matching algorithms’ switch-based approach has been developed to improve these services’ latency. GALT and matching algorithms are offloaded to the Ethernet switch with this approach. Packet forwarding and modification by reducing fat tree-based networks’ traffic are implemented on the switch. So RTI, which behaves like distributed operating system, services are offloaded to the switch hardware in many processors network. This reveals distributed operating system hardware of parallel discrete event simulation for many processors on the fat tree network.

Calculated cost functions and parallel discrete event simulator (PDES) results of the intelligent switch (IS) are compared with conventional switch (CS). Node sizes of 16, 54, and 128 and message lengths between 64B to 1024B are swept during these comparisons. Calculated cost functions show that the GALT algorithm is improved by up to 72% and the matching algorithm is improved by up to 19% with IS under no traffic conditions. GALT algorithm improvement is increased with network size by scalable IS. PDES test reveals that scenarios using GALT and matching algorithms implemented on IS improves the simulation performance by up to 53% under different traffic loads. Performance improvements are increased for low publish/subscribe ratio likewise the fine-grained applications. At last but not least, the transient message problem revealed during GALT algorithm’s optimistic synchronization implementation is solved with this study. Simulation performance is improved by removing the rollback operation. Scalable IS computes the GALT and matching algorithms in the network and also solves the transient message problem, helping to improve the performance of fine-grained parallel discrete event simulations.