At the moment I only need to deal with a couple dozen Linux images and each has different access patterns based on the application that runs in
it. I doubt Linux (SUSE) will have any of the complicated channel programs that would confuse MDC. If I use the assumption of no complicated channel programs, just straight - forward reads and writes, where would I find info on (or how would I determine) the relative performance metrics of cache read/write hits & ; misses? I have good stats on the MDC cache hit ratio and read/write ratio of each guest. If I had reasonable metrics for the relative performance of cache hits vs. misses for reads and writes I could take a decent stab at determining which Linux guests are good candidates for using MDC or not. Brian Nielsen On Thu, 13 Jul 2006 11:56:43 -0400, Bill Bitner <[EMAIL PROTECTED]> wrote: >Brian, you raise an interesting question about tuning MDC on a >per user basis vs. a system basis vs. a combination. A single >solve all algorithm would be cool, but beyond my imagination >to make it perfect. I believe there was a lot of research >originally about > >While MDC is a write-through cache, it does not automatically >insert writes into the cache if the disk location being >updated is not already in the cache. This makes it a little >more forgiving for workloads that have slightly higher >write activity. I wonder if the 80/20 rule could be >applied with something such as if my read/write ratio is >lower than "n", do not use MDC. The tricky part is >determining "n". If the ratio is 1 or lower, the data >is probably read once, write once and won't benefit from >the cache. >Part of what it depends on, is the write channel program >some are worse than others. In most cases, MDC can >determine which blocks are involved with the write I/O >and handle appropriately. There are exceptions where >complicated write channel programs confuse MDC and for >integrity reasons, it will purge more of the cache then >it needs. >Those are my few random thoughts on this. Emphasis on >random. :-) > >Bill Bitner - VM Performance Evaluation - IBM Endicott - 607-429-3286 >======================== ========================= ========================
