Message Passing for Linux Clusters with Gigabit Ethernet Mesh Connections
IPDPS '05 Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05) - Workshop 9 - Volume 10
Proceedings of the 20th annual international conference on Supercomputing
High-performance message-passing over generic Ethernet hardware with Open-MX
Parallel Computing
| Hi-index | 0.00 |
This paper proposes a high performance communication facilityusing multiple commodity network interface cards (NICs).Called PM/Ethernet-kRMA, it is NIC-hardware-independent andprovides (k)ernel-level Remote Memory Access (kRMA) on multipleNICs.The PM/Ethernet-kRMA communication protocol isprocessed on the host processor, and the protocol handler accessuser data space directly from the kernel, and then transfersthe data to network using existing network device drivers.This protocol provides one-copy communication between usermemory spaces on kernel.The PM/Ethernet-kRMA is implemented using the PM/Ethernet,one of the communication facilities of the SCore Cluster systemsoftware on Linux.The PM/Ethernet uses the Network Trunkingtechnique, which provides message communication using multipleNICs.Existing protocols, such as TCP/IP, can be used onthe PM/Ethernet-kRMA as well as the PM/Ethernet.We have evaluated the PM/Ethernet-kRMA using 2-node singleXeon 2.4GHz processor machines with three Intel PRO/ 1000XTs and one Broadcom 5701 based Gigabit Ethernet NICs oneach node.Network Trunking provides 420 MB/s of communicationbandwidth using four Gigabit Ethernet NICs.PM/Ethernet-kRMAusing four Gigabit Ethernet NICs, in contrast, provides487 MB/s of bandwidth which is 97.4% of hardware-level bandwidth (500 MB/s).