 |
The Question is:
I'm involved with a magnetic fusion experiment that
stores about 5 Mbytes of data in a burst every few
minutes up to about 200 times a day. When the data is
stored a number of processes access it right away. I am
looking for a way to make this data access quicker. I
was hoping there is a way to set up a cache that holds
maybe the last 100 Mbytes of data written to a disk in
memory. Any suggestions?
The Answer is :
RMS global buffering and XFC, or XFC, particularly given sufficient
physical memory for buffers and sufficient (though slower) disk
storage, and reasonable process quotas.
Five megabytes of cached data is well within the XFC design and
expectations, in terms of the available potential XFC cache size
and the associated cache speed.
RMS files and global buffering provides a simple, effective, and
fast way to access the data -- and provides the interlocking that
is often found necessary. Custom-written schemes generally do not
initially provide interlocking, and this omission tends to cause
application problems over time -- and a need to retrofit the locks
and the multi-processor (and potentially clustering) capabilities.
The use of RMS global buffers does assume use of RMS records and
record operations. RMS global buffers do not cache block operations.
Use of XFC is default for all virtual I/O and does cache RMS record
and block operations, as well as any $qio or $qiow using virtual
I/O operations, for all I/O operations with the upper transfer size
limit (for purposes of XFC caching) determined by the VCC_MAX_IO_SIZE
system parameter.
You will want to discuss details of the application and its requirements
with HP services folks or other knowledgeable folks specializing in
this design area. Details such as the penalty for data drops -- does
the data stream recover from transient data loss, for instance -- the
timeliness requirements for getting the data to disk, for the immediacy
of access to the data from other processes, the value of the data
itself, the responsiveness of the processor and contention requirements
and restrictions, storage and I/O hardware, and various other aspects
all affect this and most any other real-time design.
|
|
|
|
|
