On 20 March 2012 23:11, Graham Harris <[email protected]> wrote: > >> We do not seem to be getting the results we thought we would. On this >> particular LPAR, we we're running with 3CPs, but when we activated a 4th >> CP, our critical path ran longer. Overall we are pushing through more >> work, but our critical path window ran longer. We are stumped on why. >> We're starting to look at RMF data, but thought I would throw it out to >> see if anyone has any clues. >> >> > When you add extra engines, the extra MP effect means they all effectively > get a little less powerful. Depending on the machine model, this would > tend to be around a 2-3% reduction. So what you are seeing (improved > overall throughput, but extended runtimes) is not an unreasonable thing to > happen. > > Are you running with Hiperdispatch turned on? Conventional wisdom may say > its not worth using HD with so few CPs, but there may potentially be > valuable cache-locality-retention benefits with your CPU-intensive batch > jobs (depending on how well they may benefit from this, of course). > If you are running with HD off, then you are now running more concurrent > workload on 4CPs which could be polluting all your processor caches a lot > more than you were experiencing previously on 3 CPs, which could lead to > increased overheads in the OS managing higher levels of cache-misses under > the covers. > So if you are running with HD OFF, it may be worth trying with HD=ON to > see if it makes any sort of difference. > > HTH >
Having wrote all that, I just realised that with 4 CPs, then they're all likely to be present on a single book anyway (and single affinity node? Although do affinity nodes even exist if HD is off?). Hopefully 'them that know' can clarify.... May be helpful if the OP could specify what machine model was upgraded from/to (or is it actually really a bigger machine than 4 CPs, and was perhaps just an extra CP config'd on to the LPAR in question?), and whether other LPARs share the box. ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
