Infiniband and SAN performance.

Fibre Channel is the predominant storage area network (SAN) technology today. Internet small computer system interface (iSCSl) solutions have been in existence for a while, but adoption is limited and waiting for mainstream market pricing and the adoption of 10-Gigabit Ethernet. InfiniBand is a

With InfiniBand, packet delivery is handled in the hardware, not in the software. Through use of credit-based flow-control mechanisms and centralized management and monitoring of available bandwidth in the fabric, InfiniBand delivers packets between sending and receiving nodes in a lossless way.

Hardware-implemented, congestion-control mechanisms ensure that such services are guaranteed in hostile environments. Granular and hardware-level quality-of-service features ensure that storage-based I/O traffic gets the needed level of service in unified I/O usage scenarios.

InfiniBand adapter and switch solutions of 10 Gbps and 20 Gbps are available from most major server OEMs.

InfiniBand uses cost-effective and scalable remote directory memory access (RDMA) and transport offload technologies to deliver high throughput and low latency. Software stacks required for SCSI and iSCSl over RDMA programming interfaces are available in major operating systems and deliver SAN bandwidths of up to 1,500 Mbps.

InfiniBand adapters support channel I/Obased services, where each channel (along with underlying hardware resources allocated per channel) can be dedicated to separate traffic types--storage, networking, clustering and management.

InfiniBand adapters provide unified and scalable I/O connectivity on the server side, reducing cabling complexity and improving I/O utilization and cost. When scaling capacity on the SAN side, data center designers can choose either native lnfiniBand-attached storage targets or modular InfiniBand-to-Fibre Channel bridges to connect to existing Fibre Channel SANs and back-end storage. In either case, the high levels of bandwidth and reliability are maintained for SAN-based applications running on the servers.

The amount of available end-to-end SAN bandwidth is dependent on back-end storage capacity on the SAN side. When modular bridges are used, remote Fibre Channel ports can be shared, SAN capacity can be designed and scaled based on average load across servers, and SAN and server capacity can be scaled independently.

The server deployment paradigm has changed from symmetric multiprocessing servers to clusters of commodity servers, enabling improved scalability and flexibility. Clustered database offerings capitalize on those trends to offer high performance and scalable database solutions at a fraction of the cost.

With use of InfiniBand I/O, CPU utilization and latency is cut in half for server-to-server messaging, shifting the bottleneck to SAN performance, more specifically data transfer and seek times from remote disks. With InfiniBand-attached storage, and a 14-disk storage array, SAN performance of up to 640 Mbps has been achieved.

Multicore CPU adoption and the need for higher hardware resource utilization is driving enterprises to adopt server virtualization solutions. Blade servers, however, lack real estate to hold multiple I/O adapters for LAN and SAN connectivity, as required by such deployments. I/O unification becomes critical, and there is a growing need to contain the number of I/O adapters needed per physical server.

InfiniBand's I/O unification capabilities and higher throughput makes it useful for such applications, reducing costs by about 50 percent. In virtual server environments, InfiniBand has shown it can deliver native SAN performance (1,500 Mbps) from virtual machines, and scale almost linearly across multiple virtual machines.

Sujal Das is senior director of software product management at Mellanox Technologies, Santa Clara, Calif.

For more information: rsleads.com/706cn-255