Cisco InfiniBand 4x User's Guide download pdf (Page 51)

Chapter 3. InfiniBand technology 37

underlying iSCSI-like mechanism. This does not preclude emulating a Fibre Channel

connection, while enabling a much more functional mode, where applications can have

visibility to mixed SAN fabric environments without having to understand and account for the

differences between those fabrics.

3.2.4 User-level Device Access Programming Layer

The Direct Access Programming Layer (DAPL) protocol for user-space applications (uDAPL),

and for kernel mode use (kDAPL) are relatively new industry standards for high-efficiency,

low-latency server-to-server communications. User-level Device Access Programming Layer

(uDAPL) is one of the lowest-latency, highest-bandwidth, and highest-efficiency standard

protocols available on InfiniBand. MPI is currently slightly better then uDAPL, but additional

performance tuning can put uDAPL in first place. uDAPL is the preferred interface for new

development in low-latency commercial applications. Database vendors (such as Oracle®

with RAC 10i) will be using uDAPL for scale-out database clustering.

uDAPL has also been used as the underlying support for other protocol implementations.

Specifically, Scali has delivered MPI support for InfiniBand by layering their MPI stack on top

of uDAPL on InfiniBand. Scali is counting on uDAPL being a small, simple, common layer on

top of which MPI can still provide significant performance and efficiency benefits.

kDAPL is a similar interface for use from kernel mode applications or from the operating

system itself. Primary use is aimed at the file system interface. NFS/RDMA, iSER, and other

storage interfaces could go directly through kDAPL.

IT API is an industry-standard initiative (also involving IBM) working to standardize a common

API for commercial applications. At this time, uDAPL is providing this function. This effort

might, at some point, come up with another interface. However, with the momentum behind

uDAPL, it is likely that the

new interface will be based on what we have now.

3.2.5 Message Passing Interface

Message Passing Interface (MPI) is the standard protocol for scientific and technical

high-performance clustering (HPC). MPI can be used over many different cluster fabrics,

including Ethernet, Myrinet, Quadrics, and InfiniBand. MPI provides the lowest latency,

highest bandwidth, and highest efficiency of all of the standard protocols available on

InfiniBand.

Most HPC Linux environments utilize MPI protocols. Much of this market is led by work done

at national labs and research universities around the world. The MPI stack (MPICH)

developed at Ohio State University is the most commonly recommended MPI implementation

with the Topspin InfiniBand solution.

1 2 ... 46 47 48 49 50 51 52 53 54 55 56 ... 173 174

No comments

Cisco InfiniBand 4x User's Guide Page 51