Conclusions

In this paper we explore the use of remote main memory for paging. We describe our prototype implementation of a pager on top of the DEC OSF1 operating system as a device driver. No modifications were made to the kernel of the (monolithic) DEC OSF1 operating system. We run several applications to measure the performance of the system. Based on our implementation and our performance results we conclude:

• Paging to remote memory results in significant performance improvement over paging to disk. Applications that use our pager even when running on top of traditional Ethernet technology show performance improvements of up to 112% (see figure 2). Extrapolating from our results, we show that on top of a faster interconnection network even higher performance improvements are realizable!
• Paging to remote memory is an inexpensive way to let applications use more main memory that a single workstation provides. Remote memory paging provides good performance with almost no extra hardware support. The only way for magnetic disks to provide comparable performance is to use expensive disk arrays.
• The benefits of paging to remote memory will only increase with time. Current architecture trends suggest that the gap between processor and disk speed continues to widen. Disks are not expected to provide the bandwidth needed by paging unless a breakthrough in disk technology occurs. On the other hand, interconnection network bandwidth keeps increasing at a much higher rate than (single) disk bandwidth, thereby increasing the performance benefits of paging to remote memory.
• Reliability comes at little extra cost. Our parity caching method introduces at 5% more page transfers than the simple remote memory paging methods that provide no reliability: a small cost to pay for such a large benefit.

Based on our performance measurements we believe that remote memory paging is the only cost and performance effective way to execute memory-limited applications, on a network of workstations.