Re: Dbload performance

In article <1995Nov23.222255.4190@phzzzt.atww.org> mfaurot@phzzzt.atww.org (Michael Faurot) writes: >Rob McAllister (rmcall@on-ramp.ior.com) wrote: >: 1) Don't create indexes until AFTER dbload is finished. >: >: 2) Increase the number of rows in a transaction (see -n option for >: dbload). >: >: 3) Better than 2 is to turn off logging all together during the load. > >I do exactly the same thing to do a 17M row import from a 1.7G ascii >flat file. Using nawk to do some very minor conversions, I write my >output to a named pipe and then have dbload read from that named pipe. >I've been able to get between 550-600 rows into my database per second. >The ASCII rows are about 144 bytes each. The platform is a Pyramid >MIS-4S running Informix v7.01. On multi-cpu machines, I've had more fun having multiple dbloads reading multiple named pipes. I wrote a simple program that splits (round-robin) an input stream into multiple output streams feeding the dbloads. Many times a single dbload is cpu-bound vs. disk-bound. I did this on a 16 processor Sequent and was amazed at the increase. Scott