Re: KAIO ands "onstat -g iov" output

In article <932h3l$os4$1@nnrp1.deja.com>, William Rice <ricew@operamail.com> writes >I can't imagine a couple of extra AIOVP's having anything like a >significant impact on performance. > >Will >1 AIO VP's a) Use more memory - seperate process b) Use more thread scheduling time - since they get woken up if the first one is busy. c) Result in more flushes of the CPU's internal cache's since memory mapping on the CPU change more often d) Are more like to show up bugs in AIO VP Code since >1 running means the locking/latching code gets activate and this code may have bugs! e) Result in more waiting for latches/locks due to d) [Note, however that sqexplain.out files get written by AIO VPS! I think it is the MISC or ADM VP which creates the file but it then closes it and the AIO VP opens and writes to it. Yes I've been playing with truss again on Solaris 2.6, that one is not in the manual]. -- David Williams

In article <YV1H3aA8VOV6Ew70@smooth1.demon.co.uk>, David Williams <djw@smooth1.demon.co.uk> wrote: > In article <932h3l$os4$1@nnrp1.deja.com>, William Rice > <ricew@operamail.com> writes > >I can't imagine a couple of extra AIOVP's having anything like a > >significant impact on performance. > > > >Will Do note I said significant impact... > > >1 AIO VP's > > a) Use more memory - seperate process Which will get swapped out if it's never used, maybe I am spoiled, but I am not to concerned about the memory taken up by a couple of extra aiovp's. > b) Use more thread scheduling time - since they get woken up > if the first one is busy. But we want good throughput anyway, if it get's woken up it was needed. > c) Result in more flushes of the CPU's internal cache's > since memory mapping on the CPU change more often You lost me here. If you are saying there will be more context switches on the CPU then if it is active often enough to cause a significant number of context switches it was probably needed for good i/o throughput. > d) Are more like to show up bugs in AIO VP Code since >1 running > means the locking/latching code gets activate and this code > may have bugs! > e) Result in more waiting for latches/locks due to d) I have seen bugs with KAIO, not with AIO, that is just me though Once again, if I am writing to enough devices that locking and latching overhead are going to be a significant slowdown, I probably have enough write activity to buffered devices to warrant multiple aiovp's. I have been wrong many times before, but unless the number of aiovp's is somehow causing disk thrashing, I can't see +- a few having a significant impact. Will > > [Note, however that sqexplain.out files get written by AIO VPS! > I think it is the MISC or ADM VP which creates the file but it > then closes it and the AIO VP opens and writes to it. Yes I've > been playing with truss again on Solaris 2.6, that one is not > in the manual]. > > -- > David Williams > Sent via Deja.com http://www.deja.com/