HP-PIAF I/O Bottlenecks

The current disks in PIAF achieve only 4.5 MB/sec (RAID 0) and 6.5 MB/sec (RAID3)
- These are aggregates across the RAID arrays: individual disks run at 1.5 MB/sec
Each Slave usually is reading the nTuple from another Slave’s disk
- The Slave machine hosting the disk containing the nTuple must cope with 7 concurrent accesses to the file via the network, as well as it’s own
- If the file were to be striped across N seperate disk channels, then the aggregate access rate could reach N*6.5 MB/sec. Except ...
Memory to memory transfers across FDDI in HP PIAF achieve only 7 MB/sec
- The HP FDDI cannot make more than 50% of the available bandwidth, but in any case ...
The default NFS blocksize is too small for efficient transfer
- The maximum appears to be 8096
- Which limits the throughput (for a file already in the NFS cache) to around 7 MB/sec