Re: KAIO vps, and bufwaits

Tony Tyrwhitt-Drake wrote: > Neil Truby wrote in message <34DB88BE.56A3@netcomuk.co.uk>... > >Does anyone know how to set up extra KAIO vps? > The KAIO bits are connected to the CPU vp bits. If you add another CPU vp > you will appear to get an extra "KAIO vp" although I think you'll find it is > actually part of the CPU vp. > >Also, despite having 200M of BUFFERS, onstat -p is still showing some > >bufwaits. The read cache rate is above 99%, but the write remains > >around 87% - the same as on the Sparc when BUFFERS was set to 50M. The > >new system isn't swapping and hardly pages; do the bufwaits suggest a > >further increase in BUFFERS is warranted? > The bufwaits may be there as a result of checkpoint activity. Your write > cache rate will depend on what program code you are running. If applications > are doing mostly updates the cache rate will climb and with mostly inserts > it will fall. You can't do a lot about it so I wouldn't worry too much. > If you have the physical memory, more BUFFERS is hardly ever a bad > idea...but if the thing is performing OK and the punters are happy, leave it > as it is. I'll just add one more thing to this discussion since I agree with Tony and Neil, your cache %ages look OK unless we know more about the mix of inserts -vs- deletes and updates. Any write cache rate over ~75% is good for an insert intensive server. Anyway, if you have LRUS set to either 64 or 96 you may be hitting an as yet unverified bug I have been tracking having to do with the hash algorithm used to select the next LRU to pick to get a clean buffer when the currently select LRU is locked by another session. If this is the case changing LRUS to either 32 or 128 is known to reduce BUFWAITS substantially. The bug shows up at LRUS = 64 or 96 and maybe other values as well. If you feel 32 is too few and 128 too many you might try some prime number in between, I only know for sure that these two values are OK. Please let me know if that was the problem and how you solved it, I need fuel to force R&D to look at this one seriously. Also be sure to adjust CLEANERS appropriately when you adjust LRUS. Art S. Kagel