The HiPPI Solution

Measurements of HiPPI memory to memory transfers between two PIAF machines indicate a throughput of around 15 MB/sec
- The theoretical HiPPI point to point bandwidth is much higher ...
Measurements of disk to memory rate between these two PIAF nodes shows a throughput of 6.5 MB/sec
- i.e. as if the disk were local ...
To remove the disk I/O bottleneck, we must implement PIAF file striping
- With Timo’s Stager
- The bottleneck is then the network layer, so ...
We should put in HiPPI for an aggregate nTuple access rate of around 50 MB/sec (and serial HiPPI for transfers between PIAF and CORE)
- The bottlneck is then the NFS UDP packet size, so ...
We implement RFIO in PIAF, instead of NFS, with big blocks.