Andy Polyakov wrote:
Another thing I would do is to run "vmstat 1" and see if the periods of bad response match a very high rate of irq or ctx. Fuzzy hardware indicated, but I would love for him to try a vendor kernel as well.
OK, requestor has reported that readcd performs at >2-3x accompanied by irresonsiveness of the system. This *does* appear as lack of DMA and it agrees pretty well with those reports from NEC users I mentioned yesterday. Yes, DMA must be off (or not being used on that particular data flowpath), *but* I don't think it's the factor resposible for further <1/5x performance loss, from >2x to 0.4x that is. In other words lack of DMA is the *least* problem and search for limiting factor may *not* stop there. A.
-- E. Robert Bogusta It seemed like a good idea at the time
