We study a simplified version of the Tivoli Data Exchange [20] system for replicating data across large numbers of hosts. This system distributes data and programs across thousands of client machines using a hierarchy of replication servers. Both non-interference and good throughput are important metrics. In particular, these data transfers should not interfere with interactive use of target machines. And because transfers may be large, may be time critical, and must go to a large number of clients using a modest number of simultaneous connections, each data transfer should complete as quickly as possible. The system currently uses two parameters at each replication server to tune the balance between non-interference and throughput. One parameter throttles the maximum rate that the server will send a single client; the other throttles the maximum total rate across all clients.
Choosing these rate limiting parameters requires some knowledge of network topology and may have to choose between overwhelming slow clients and slowing fast clients (e.g., distributing a 300MB Office application suite would nearly a day if throttled to use less than half a 56.6Kb/s modem). One could imagine a more complex system that allows the maximum bandwidth to be specified on a per-client basis, but such a system would be complex to configure and maintain.
 |
Nice can provide an attractive self-tuning abstraction. Using it, a sender can just send at the maximum speed allowed by the connection. We report preliminary results using a standalone server and client. The server and clients are the same as in the Internet measurements described in Section 5. We initiate large transfers from the server and during that transfer measure the ping round trip time between the client and the server. When running Reno, we vary the client throttle parameter and leave the total server bandwidth limit to an effectively infinite value. When running Nice, we set both the client and server bandwidth limits to effectively infinite values.
Figure 11 shows a plot of ping latencies (representative of interference) as a function of the completion time of transfers to clients over different networks. With Reno, completion times decrease with increasing throttle rates but increase ping latencies as well. Furthermore, the optimal rates vary widely across different networks. However Nice picks sending rates for each connection without the need for manual tuning that achieve minimum transfer times while maintaining acceptable ping latencies in all cases.