[robespierre]

The PCMCIA card seems to be the hard limit on bandwidth. I tried block sizes up to 4MB with no increase in performance. Larger blocks probably would work fine, but I like to stick with 512 byte block sizes 1) because I'm never fully sure how some of the archaic hardware will handle big blocks and 2) I regularly use SCIS2SD devices, which have a default block size of 512 bytes, so I don't have to do mental math to target the right sector.

PCMCIA I/O interfaces operate at roughly 3.3 MHz on a 16-bit data bus, for a bandwidth of 6.6 MB/s without pipelining. The interface is based on the AT bus with relaxed timing requirements (the AT bus ran originally at the CPU's clock frequency, but later on was commonly used at anywhere from 7 to 8.3 MHz). A SCSI PC Card should be able to saturate a 5 MB/s SCSI bus.

You didn't reply to or acknowledge my comment about cooked devices. The block size used in a write request can only influence the actual device transfer if the raw device is used (with an "r" in its file name). The cooked device always stores requests in the file cache and uses a scheduling policy to choose when to actually issue them and what block sizes to use. It's like the difference between writing to a TCP stream and sending a UDP datagram, only one offers control over the packet size to applications.