Sean Liu wrote: > Now - let's take a look from another angle: > a. When people take a look at vmstat output, they don't expect > an exact number of bytes of free memory, they need a rough number > however. So some reasonably accurate number would be good enough.
Richard Elling <richard.ell...@gmail.com> replied: | good enough may be reasonably achievable, just subtract arcstats:c_min [big snip] | We lost the clue with the invention of demand paged virtual memory | 30 years ago. Actually Sean's asking for a clue about the *virtual* memory behavior. In SunOS 3 and 4, free memory told the observer how much virtual memory was available, and was a linear metric: if you had 20% free, you had twice what you had when the report was 10% free. In SunOS 5, we lost this kind of metric. Free memory went to zero soon after the systems started. I was working at Opcom (Operation Commitment, the 5.0 lab rats) and the complains from the customers came in loud and fast. They didn't enjoy being told that there was no direct measure of memory. Then came the scan rate kludge: if the scan rate jumped up, you had run out of memory. It was a non-linear measure, roughly exponential with the decrease in free pages, and it only told you *after* you got in trouble. Customers usually want to but more memory before they get into trouble, so this was particularly disappointing to our large, business-critical customer base, the very people we most wanted to support. For them, it was have a unexpected performance hit at the busiest time of the year, diagnose it, call Sun and then try to get a memory board bought and into place before the company went up in a little puff of greasy black smoke (;-) Most of us used scan rate and IOPS on the swap partition or disk as a guesstimator, the latter being a bit more linear. I'm a capacity planner, so this was a particular PITA for me: it meant I couldn't credibly do predictions of future memory usage. So I got slagged by my customers... Then Richard McDougall introduced Priority Paging, and the free memory metric went back to being useful and linear once more. I was more than a tiny bit relieved. The conclusion to this history? We want something that credibly tells us that applications are running out of memory, and we'd prefer it tell us before the app crashes, and be linear so we can predict at least a little ahead. If (free - arcstats:c_min) fits this bill, then it's cool! --dave [ Hey, is that Sean from was used to be SSHA?] -- David Collier-Brown | Always do right. This will gratify Sun Microsystems, Toronto | some people and astonish the rest dav...@sun.com | -- Mark Twain cell: (647) 833-9377, home (416) 223-8968, bridge (877) 385-4099 code 506 9191# _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
