Chapter 2 HPSS Planning138 September 2002 HPSS Installation GuideRelease 4.5, Revision 2application, a number of the disks can be grouped together in a striped Storage Class to allow eachdisk to transfer data in parallel to achieve improved data transfer rates. If after forming the stripegroup, the I/O or processor bandwidth of a single machine becomes the limiting factor, the devicescan be distributed among a number of machines, alleviating the limitation of a single machine.If the client machine or single network between the client and HPSS becomes the limiting factor,HPSS supports transferring data to or from multiple client machines, potentially using multiplephysical networks, in parallel, to bypass those potential bottlenecks.During system planning, consideration should be given to the number and data rates of thedevices, machine I/O bandwidth, network bandwidth, and client machine bandwidth to attemptto determine a configuration that will maximize HPSS performance given an anticipated clientwork load.2.11.5 ConfigurationThe configuration of the HPSS storage resources (see Section 2.8.6) is also an important factor inoverall HPSS performance, as well as how well the configuration of those resources matches theclient data access patterns.For example, if a site provides access to standard FTP clients and allows those clients to write datadirectly to tape, the buffer size used by the FTP server and the virtual volume block size defined forthe Storage Class being written to will have a significant impact. If the buffer size used by the FTPserver is not a multiple of the virtual volume block size, each buffer written will result in a distinctstorage segment on the tape. This will cause additional metadata to be stored in the system andextra synchronization processing of the tape. (If the buffer size is a multiple of the virtual volumeblock size, each write will continue to append to the same storage segment as the previous write.This will continue until the final write for the file, which will usually end the segment, thusreducing the amount of metadata generated and media processing.)2.11.6 FTP/PFTPData transfers performed using the standard FTP interface are primarily affected by the buffer sizeused by the FTP Daemon. The buffer size can be configured as described in Section 7.3: FTP DaemonConfiguration (page 422). It should be a multiple of the storage segment size, if possible. If not, itshould be, at least, a multiple of the virtual volume block size. If the buffer size is too small, the FTPDaemon will need to issue a large number of individual read or write requests; however, if thebuffer size is too large, the FTP Daemon will require a large amount of memory, which may causeadditional paging activity on the system.The size of the FTP Daemon buffer is extremely important if the FTP clients write files directly to atape storage class, as described in Section 2.11.5.Parallel FTP (PFTP) can be used to move data using TCP/IP.Note that the PFTP data transfer commands (e.g., pput and pget) are not influenced by the FTPDaemon buffer size because the data flows directly between the client and Movers.Note that PFTP clients that use the standard FTP data transfer commands (e.g., put and get) havethe same performance considerations as standard FTP clients as described in Section 2.11.6.
