Writing to HDF5 file timing is erratic

Question

Dear Community,
I am trying to acquire and stream 10-15 megabytes / sec from a bunch of 4472 cards to an HDF5 file and I am having a problem that every once in a while, the H5Fwrite vi takes a lot of CPU time (3-10 seconds) to write a chunk to disk, during which time the backlog from the acquisition system gets too big and overflows the buffer.
I have tried working around this by creating two loops, one of which reads data from the 4472s and puts it on a queue, while the other dequeues the data and writes it to disk.
But I'm also wondering whether there's a way to control when the HDF5 vi's actually write to disk, and whether there are some parameters I can tune in the HDF5 libraries, such as block sizes and cache sizes
, so that writing to disk is fast and not so "lumpy."
I've tried using the HDF5 library functions h5Pset_cache() and h5Pset_sieve_buf_size() to change the sizes of various buffers but it doesn't seem to change anything.
Or maybe I should just use the SFP libraries to do this instead instead of rolling my own HDF5-based vi's?
Thanks,
Cas
Casimir Wierzynski
Graduate Student, Computation and Neural Systems
California Institute of Technology, MC 139-74
Pasadena, CA 91125

Answer 1

What hardware are you using? A 2000 era computer (Pentium III 650) should be able to sustain 10MBytes/sec or better until you run out of disk space. A 2004 era computer (Pentium 4 2.4HT) should be able to do twice that. Both are disk speed limited. Are you using a PXI controller or a desktop machine? PXI controllers use laptop hard drives to keep power consumption down. Laptop hard drives are considerably slower than those in a typical desktop machine.
Some general tips which may help you.
1) Make sure all other processes are turned off. This includes screensavers, the Microsoft indexing service, virus checkers, etc. This can be tough on an NT, 2000, or XP machine, but is absolutely vital if you want to sustain high speeds.
2) Tune your HDF5 chunk size appropriately. For Windows based systems, the optimum is 65,000 points (just a bit under the 65,525 16-bit boundary). If this chunk size gets anywhere near the 1MByte size of the normal HDF5 cache, your write speeds will take a huge hit.
3) Tune your fetch size from the 4472s for maximum performance. I would expect this to be in the low 100s of thousands of points (that is where it is for NI high-speed digitizers). Unless you increase the HDF5 buffer size from 1MByte, you probably should keep this below about 350,000 points to avoid paging problems. Just fetch the large buffer and write the whole thing at once. The HDF5 software will chunk it to disk for you.
4) Make sure you serialize calls to the HDF5 driver. HDF5 is NOT threadsafe under Windows (it is under Linux). This could cause serious data corruption and other wierd problems if you are taking data from multiple boards and writing to the same file in multiple loops.
5) Check your disks to make sure you do not have a hardware problem. A dying hard drive will produce symptoms similar to these. You can get utilities from the major disk manufacturers.
If you tried all these things and are still having problems, please reply with more information. You should be able to succeed. You may also want to check out NI-HWS, available on the latest driver CD. It uses the same HDF5 format as the SFP routines, but is far easier to use.
This account is no longer active. Contact ShadesOfGray for current posts and information.