On Wednesday June 9, bugzilla@xxxxxxxxxxxxxxxx wrote: > You said: > "Now consider RAID5. Here, with a hardware controller all of the data is > written to the RAID card which in turn calculates parity and stripes the > data over the disks. With software RAID, the software calculates parity and > writes the data across all the mirrored drives. The only additional bus > traffic for software RAID is the parity data." > > I believe this is wrong: > "The only additional bus traffic for software RAID is the parity data." > > It is true if 100% of a stripe is being changed/written. > If you update less than 100% of a stripe the software RAID must read the > full blocks being changed and the parity block. Factor out the old data > from the parity then compute a new parity. Then write the new blocks. > > Example: > Your array will have 6 disks. You don't state your block size, so > let's assume 64K. Your stripe size will be 5*64K or 320K. Now if you were > to write 1 byte to your array this is what will happen: > Read the 64K block that contains the 1 byte. > Read the 64K parity block. > Factor out the 64K data block from the parity block. > Merge your 1 byte into the 64K data block. > Compute a new 64K parity block. > Write the new 64K block that contains your 1 byte. > Write the 64K parity block. This is mostly correct, except that it won't be a 64k block. It will normally be a 4k block. Your chunksize is irrelevant. In 2.6, md will do a PAGE_SIZE read/write, which is 4k on x86. In 2.4, md will do read/writes that match the filesystem blocksize, which is most often 4k these days. > > As you can see, your 1 byte require reading 128K from 2 different disks, and > then writing 128K to the same 2 disks. So that's 8k, twice. > > I don't know how md really does this. I have not looked at the code. > Another choice would be to read 100% of the strip, apply your updates (1 > byte in my example), then compute the parity, then write the changed > blocks. md sometimes does a "read-modify-write" cycle like your first example, and sometimes does a "reconstruct-write" cycle like your second example. It chooses the option that generates the fewest IO requests. NeilBrown - 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
