Re: KAIO perf on HPUX

Posted in 1997

David Williams — Fri, 1 Aug 1997 23:05:36 +0100 — source: @@NL@@>As part of our upgrade to 7.23 I switched on Kernel Asynchronous@@NL@@>IO as this had been recommended to me by an Informix consultant@@NL@@>(whose name I don�t remember - shame). I have read the Admin.@@NL@@>guide on the subject of KAIO and it seemed a sensible thing to@@NL@@>do. I believe I understand what is supposed to happen, i.e. IO is@@NL@@>handled by the cpu vp (of which we have one to match the single@@NL@@>processor that the instance is configured for) and in doing so@@NL@@>avoids the overhead of the context switch to an AIO vp.@@NL@@>Unfortunately it had a very detrimental effect on performance. A@@NL@@>10 minute query took over half an hour. Clearly there is more to@@NL@@>the mechanics of KAIO than I understand.@@NL@@@@NL@@ Check your UNIX kernel parameters for any parameters related to@@NL@@ KAIO, possibly there is a limit of the number of KAIO requests@@NL@@ active at any one time, try uping that limit.@@NL@@@@NL@@ Try adding a second CPU VP. Possibly the one CPU is a bottleneck,@@NL@@ adding KAIO means the CPU VPs are in charge of sending I/O requests@@NL@@ to the kernel. Hence more load on the CPU VPs, more reason to try@@NL@@ adding a second one.@@NL@@@@NL@@@@NL@@@@NL@@@@NL@@>\

In article <5rt9pb$aq6$1@mhadf.production.compuserve.com>, Martyn Hodgson <101330.3717@CompuServe.COM> writes >Hi, >We are running Online version 7.23 on a 2 processor HP machine >(HP-UX version 10.20). The Online Instance is configured as if >the server has just a single processor, as Informix recommend >this is more efficient (I believe that the theory is that a >single cpu avoids having to some locking and synchronisation and >is therefore faster than a twin processor machine). All of our Not always, it depends upon the overhead of locking, sychronisation and cpu switching vs the gain by using a second CPU. The correct recommendation is to TRY it both ways and see which is faster. Which one is faster depends on the exact hardware you are using, the verison of the operating system (semaphore efficiency) as well as the exact SQL being run and what else is running on the machine. there are so many variables that the only real answer is to try 2 CPU VPs and see if it is faster. >data is held in raw chunks. The main database is in a single >dbspaces of its own (3 raw chunks, no fragmentation). It is a >read only database, so the database is Unlogged. >\