While the focus of architectural considerations were on processor speeds and memory bandwidth, another area of importance is input and output mechanisms to secondary storage and networks. It was recognized that I/O would scale less than linearly with system performance and might not prove as difficult to achieve as some of the other challenging system requirements. Nonetheless, given recent history with I/O bottlenecks on current MPP systems, the issue warrants a deeper examination. For example, just filling the main memory of such a computer imposes a substantial burden on the I/O subsystem. Suppose one wants to checkpoint the memory contents just once an hour and the memory comprises approximately 30 terabytes as proposed by the Applications and Architectures working groups. This alone would impose an average demand of 10 gigabytes per second bandwidth. But, in fact, one would not want this overhead to be more than a small fraction of the overall system usage, or about five percent. Then the required I/O bandwidth increases to about 200 gigabytes per second. Today's disk drives might provide 10 Mbits per second peak. To absorb the checkpointing data rate on such a system would require on the order of 100,000 disk drives. This is an imposing number and makes clear that PetaFLOPS systems as being envisioned require detailed consideration of the implications for I/O and secondary storage support infrastructure. Moreover, the problem is exacerbated if the main memory is actually in the range of a petabyte as some futuristic applications were thought to require, increasing demand by more than another order of magnitude.