> While it wont explain the drop from 55-44Mbyte/second > I think ~50MByte/second is pretty good considering such a latency-bound > test such as single threaded dd writing one 1024byte block at a time. > > I would either run dd with a much much larger blocksize, or with many > concurrent dd in parallell > in order to reduce the cost-per-byte-in-latency to get throughput > numbers instead of latency-bound numbers. you can also run sgp_dd which uses POSIX thread (up to 16). |