Issues in network class loading

 

We identified several design issues that must be addressed in any network class loading technique. This section discusses these design issues and some of the tradeoffs they imply.

One network connection vs. multiple.

When the client requests class files and resources from the server, it could use a separate network connection for each file, it could use a single persistent network connection for all files, or it could use a fixed number of persistent connections. Using a single persistent connection is generally the best approach. Using separate connections for each file is generally not a good idea since setting up a network connection is expensive. It is not clear that there would be any advantage to using multiple persistent connections, since in class loading there is no obvious advantage to downloading files in parallel.

Transfer granularity: individual files vs. many.

The ability to send individual files is desirable because it allows the files sent by the server to exactly match those needed by the client. As the granularity of the transfer units increases, the probability that unneeded files might be sent increases, as does the probability that files will be sent in an order different from the request order. However, small transfer granularity increases the likelihood of of delays due to network latency, because the client must send requests to the server more frequently.

Using compression vs. not.

Compressing class files and resources is desirable when network bandwidth is limited, since it results in transferring fewer bytes over the network. However, the benefit of transferring fewer bytes must be weighed against the runtime cost of performing the compression and decompression.

Compressing individual files vs. multiple.

Applying compression to multiple files is desirable since it offers more opportunities for sharing redundant information between files, thus increasing the compression ratio.

Performing compression on-line vs. off-line.

Performing compression off-line is desirable because it reduces the amount of computation required on the server.

Pre-sending/prefetching vs. not.

When classes and resources are transferred on-demand, the client must send requests for the files it needs to the server. If files are requested only at the precise point they are needed, then client must pay the cost of network latency for each request. This latency cost can be reduced or eliminated if files are either pre-sent or prefetched prior to the point when they are needed by the client. However, care must be taken that files are not pre-sent or prefetched unnecessarily.