Conclusions

 

In this paper we described how to use several workstations in a NOW to provide fast and reliable access to stable storage. Our approach consists of using network main memory to avoid synchronous disk I/O as much as possible. By using data replication and redundant power supplies we increase the reliability of remote main memory, and use it as a short-term non-volatile storage medium.

We have implemented our approach within the EXODUS storage manager and the RVM recoverable virtual memory system. We have experimented with both systems running on top of a Network of Workstations connected with a variety of interconnection networks ranging from the traditional Ethernet, to the high-speed Memory Channel. Based on our implementation experience and performance results we conclude:

• Our approach can be easily incorporated in existing database systems. We have implemented our approach on top of two different transaction based systems. Each implementation took only a few weeks of programming, and consists completely of user-level software.
• RRVM provides significant performance improvements over RVM, even on top of Ethernet interconnection networks. Our results suggest that for small transactions, RRVM on top of Ethernet is able to sustain 70-500 transactions per second, depending on the transaction access patterns. At the same time, unmodified RVM sustains 30-50 transactions per second - an order of magnitude less. Most of the benefits of our approach are realized on top of the high-speed Memory Channel interconnection network, where RRVM manages to sustain a little more than 2,500 (short) transactions per second.
• RRVM is able to sustain several clients on top of the same interconnection network. Our results suggest that even when four RRVM systems operate on top of the same Ethernet network, their performance is 3-4 times better than the performance of RVM (for small transaction size). The performance of the same clients on top of FDDI is 5-7 times better than the performance of RVM.
• The performance benefits of our approach will increase with time. The performance of our approach is highly dependent on the latency and the bandwidth of the interconnection network on top of which it runs. Since the latency and the bandwidth of interconnection networks quickly improve with time, we expect the performance benefits of RRVM to improve at similar rates, at least in the foreseeable future.

Based on our experiments, we believe that remote memory (as demonstrated through RRVM and REX) is a viable alternative to synchronous disk I/O and should be considered seriously for implementation in databases and transaction-based systems in general.