Chapter 2 HPSS Planning
60 September 2002 HPSS Installation Guide
Release 4.5, Revision 2
a stateless protocol. This allows use of a connectionless networking transport protocol (UDP) that
requires much less overhead than the more robust TCP. As a result, client systems must time out
requeststo servers and retry requests that have timed out before a response is received. Client time-
out values and retransmission limits are specified when a remote file system is mounted on the
client system.
Thetwo main advantages of using NFS instead of a utility like FTP are (1) files can be accessed and
managed through standard system mechanisms without calling a special program or library to
translate commands, and (2) problems associated with producing multiple copies of files can be
eliminated because files can remain on the NFS server. The primary disadvantages of NFS are the
2GB file size limitations of the Version 2 protocol, the fact that UDP does not provide data integrity
capabilities, and the data transfer performance due to the limitation of sending data via the RPC
mechanism.In general, NFS should not be the interface of choice for large HPSS data transfers. NFS
is recommended for enabling functionality not provided through other interfaces available to the
client system.
TheHPSS NFS interface does not support Access Control Lists (ACLs), so don’t attempt to use
them with NFS-exported portions of the HPSS name space.
Because of the distributed nature of HPSS and the potential for data being stored on tertiary
storage, the time required to complete an NFS request may be greater than the time required for
non-HPSSNFS servers. The HPSS NFS server implements caching mechanisms to minimize these
delays, but time-out values (timeo option) and retransmission limits (retrans option) should be
adjustedaccordingly. A time-out value of no less than 10 and a transmission limit of no less than 3
(the default) are recommended. The values oftimeo and retrans should be coordinated carefully
with the daemon’s disk and memory cache configuration parameters, in particular, the thread
intervaland touch interval. The larger these values are, the larger the timeo and retrans times will
need to be to avoid timeouts.
Refer to theHPSS User ’s Guidefor details of the NFS interface.
2.5.6 MPI-IO API
The HPSS MPI-IO API provides access to the HPSS file system through the interfaces defined by
the MPI-2 standard (MPI-2: Extensions to the Message-Passing Interface, July, 1997).
The MPI-IO API is layered on top of a host MPI library. The characteristics of a specific host MPI
aredesignated through the include/mpio_MPI_config.h, which is generated at HPSS creation time
fromthe MPIO_MPI setting in the Makefile.macros. The configuration for MPI-IO is described in
Section 7.5:MPI-IO API Configuration on page 434
The host MPI library must support multithreading. Specifically, it must permit multiple threads
within a process to issue MPI calls concurrently, subject to the limitations described in the MPI-2
standard.
The threads used by MPI-IO must be compatible with the HPSS CLAPI or NDAPI threads use.
Threaded applications must be loaded with the appropriate threads libraries.
This raises some thread-safety issues with the Sun and MPICH hosts. Neither of these host MPIs
supportmultithreading, per se. They are in conformance with the MPI-1 standard which prescribes