Hello, Dan. I've looked through your patches, and have some suggestions about write operations processing. In the current implementation of the Raid5 driver the RMW operation won't begin until old blocks in the stripe cache, which are to be rewritten, become UPTODATE. But if you have dedicated h/w DMA engines, then, while an IOC(input/output contoller) performs transmition of the old strip data from the disk to the stripe cache, it may make sense to start a DMA engine, which will transmit new strip data from the bio requested to write. So, when an IOC operation complete, we'll already have all necessary data to compute new parity value. 1) For the current implementation: Trmw = Tioc1 + Txor1 + Tdma + Txor2 + Tioc2, where Tioc1 is the time it takes to update stripe cache with old data, Txor1 is the time it takes to substract old data from old parity value, Tdma is the time it takes to update strip with new data, Txor2 is the time it takes to compute new parity, and Tioc2 is the time it takes to transfer updated data to disks. So, Trmw = 2*Tioc + 2*Txor + Tdma 2) If copying old and new data to stripe cache is performed simultaneously, then time to complete the whole RMW operation will take: T'rmw = max(Tioc1, Tdma) + 2*Txor + Tioc2, where Tioc1 is the time it takes to update stripe cache with old data, Tdma is the time it takes to update strip with new data, 2*Txor is the time it takes to compute new parity, and Tioc2 is the time it takes to transfer updated data to disks. So, T'rmw = 2*Tioc + 2*Txor. (in any case, i think that Tioc > Tdma, because Tioc corresponds to the time spent reading from disk, and Tdma corresponds to operations with SDRAM, which are faster). Also, 2*Txor for (2) is less then 2*Txor for (1), because in (2) approach we have to prepare XOR engine descriptors only once, but in the (1) approach - twice. Does it make sense to revise your Raid5 driver implementaion to allow IOC and DMA to have separate destination buffers? That is, some kind of a stripe shadow. IOC will copy to the regular buffer in the stripe cache, DMA - to the shadow one. Regards, Yuri. - To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
