On Fri, Nov 10, 2017 at 4:43 PM, Alex Rousskov <rousskov@xxxxxxxxxxxxxxxxxxxxxxx> wrote: > On 11/10/2017 06:16 AM, Bike dernikov1 wrote: > >> So you suggest that we totally disable disk swap (or just for >> debuging) ? > > I would aim for totally disabling disk swap, but getting to that > configuration is easier if you keep swap enabled and consider any > non-trivial swap use as a problem that you need to fix. After all > non-trivial swap uses are eliminated, it does not really matter much > whether you keep swap enabled or not! That is also my goal. I have other scenarios in mind. We have free older servers IBM X3550 (8 cores), which can be upgrade to 48GB ram, (from other old servers) On new ones,we have only 24GB on primary, and 16GB ram on second server. For disabling swap, X3550 servers would be better, but they have only 8 cores. That could be problem, but for now we have only 0.5 load on 12/HT 24 cores. I know that core per core new cpu-s can process 4x-8x more, so more testing :) > Please note that this suggestion is specific to performance-sensitive > Squid servers -- many other servers have very legitimate reasons to use > swap. YMMV. > Oracle can be good example. Yes, we have one Oracle server, swap killed him, users was wild, load average over 10 continuosly (8 core), 16GB mem. We upgraded ram to 48GB. Now, max load is at 4-5 in peeks, 1-2 during day. Swap enabled :) it use 9GB, but use all ram :). Users don't call after upgrade. > >> That setting on production can be disaster. > > Squid swapping in production is an arguably worse disaster, as you have > learned. In many cases, it is better to deal with a lack of swap than to > rely on swap's magical effects that most humans poorly understand. YMMV. In this scenario, swap is backup cache (as I understand) ? As TIER in SAN-s. Swap could be used to translate back data to mem if used, but it stays on disk and purge after some time if not used ? Or I am in delusion ? >> We will try, we reduced number of auth helper, prolong ldap caching. >> Little steps but at last every bit counts. We reduced kerberos helpers to 100 on Friday, 20 of those 100 currently active, 13 of those have requests/replies different than zero. Now reduced to 50. No problem appear over weekend. cache_mem 14G swappines = 60 ldap cache set to 24h Mem stats: total used free shared buff/cache available Mem: 24101 16032 1488 156 6580 7497 Swap: 24561 1907 22654 > ... unless those bits are actually hurting. You may need to understand > your server memory usage better to reduce the time spent on trying > random changes. For example, if auth helpers are close to being > overloaded, reducing their number may make things worse. No overload after changes. When we saw that we don't use so much helpers it was logicaly to reduce that number in little steps. I agree, random testing can be painfully, on production even more. Can you recommend best way to analyze memory, (except free -m, top/htop). Squid-internal-mgr/mem have nice detail, i will start there with. Do you have better way ? > Alex. Thanks for help, _______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx http://lists.squid-cache.org/listinfo/squid-users