Our next set of experiments explores the network load that our RRVM system imposes. Our previous experiments suggest that on top of a lightly-loaded network, RRVM performs much better than traditional RVM. What is interesting to explore now, is what will be the performance of RRVM if several instances of it run concurrently and put significant pressure on the interconnection network. Will the network be able to handle the increased communication demands, or will it collapse? To answer this question we constructed the following experiment:
We create several instances of RRVM clients. For each client, there is also an RRVM server. All clients and all servers run on different workstations. All workstations are connected to the same interconnection network (Ethernet or FDDI). We progressively increase the number of client/server pairs participating in the experiment and measure the transactions per second each RRVM system (client/server pair) is able to sustain. Each RRVM system executes the experiment described in section 3.2.
In our experiments the log size was set to 8 Mbytes and the I/O block size to 32 Bytes. Figure 5 plots the number of transactions per second for each RRVM system (client-server pair) as a function of the number of workstations participating in the experiment. First of all, we see that the performance of the unmodified RVM system stays the same independent of how many workstations participate in the experiment. This is as expected, since RVM makes all traffic to its local disk, and does not put any network load. We also notice that the performance of RRVM-ETHERNET decreases with the number of workstations. If there is only one pair of workstations in the network, RRVM-ETHERNET sustains close to 500 transactions per second, while when four pairs of workstations participate in the experiment, it sustains close to 150 transactions per second. Even so, its performance is three times better than the performance of unmodified RVM which sustains less than 50 transactions per second. Finally, we see that the performance of RRVM-FDDI is practically constant even when eight workstations participate in the experiment. The reason is simple: FDDI has ten time more throughput than ETHERNET, and can sustain several heavily-communicating workstations. Furthermore, under heavy load, Ethernet may suffer from increased number of collisions that may lead to throughput collapse.
Figure 6 plots the results of the same experiment for transaction size equal to 2 Kbytes. Again, we observe that the performance of RRVM-FDDI stays practically the same, independent of the communicating workstations. Although the performance of RRVM-ETHERNET decreases with the number of participating workstations, it is much better than the performance of unmodified RVM system.
Figure 5: Network Load: Performance of RVM as a function of the network load - sequential accesses - all servers and all clients run on different workstations - I/O block size = 32 bytes.
Figure 6: Network Load: Performance of RVM as a function of the network load - sequential accesses - all servers and all clients run on different workstations - I/O block size = 2 Kbytes.