See ikegami's improvements above. Also my comments about using sysread rather than read which still seems to give a substantial performance improvement on my system at least.

I don't think there is much that can be done about the performance. Increasing the read chunk size probably won't benefit much as you are going to be limited by whatever buffers the system allocates to the pipe--seems to be about 4k on my system.

One thing that may improve it, even though it is counter intuative, is to insert a brief sleep after each read in the loop. Especially if the read did not return a full buffer.

If the producing process is slightly slow, then attempting to read again too quickly is pointless, as there may be nothing, or less than a full buffer load available to read, and you could end up reading a few bytes each time with a task switch required in between to permit the producer to produce some more.

By adding a short sleep, even a sleep 0; may be enough, if a read fails to fill the buffer, could improve throughput markedly. Something to experiment with on the target system and producer program.