It is strange that memPoolAlloc is less than memPoolFree; might be another overflow case. 2010/10/21 Kaiwang Chen <kaiwang.chen@xxxxxxxxx>: > Amos, > > The cache size is configured around 128GB, as reported by mgr:config: > cache_dir aufs /export/squid/cache 131072 32 256 IOEngine=DiskThreads > cache_swap_low 90 > cache_swap_high 95 > cache_mem -1073741824 bytes > memory_replacement_policy lru > cache_replacement_policy lru > memory_pools on > memory_pools_limit 5242880 bytes > > In a VIRT=11.5g RES=11g instance, mgr:info reports something like: > Process Data Segment Size via sbrk(): 1020120 KB > Maximum Resident Size: 47922032 KB > Storage Swap size: 34233780 KB > Storage Swap capacity: 25.5% used, 74.5% free > Storage Mem size: 3375668 KB > Storage Mem capacity: 46.4% used, 53.6% free > memPoolAlloc calls: 21345922346 > memPoolFree calls: 21401173704 > Internal Data Structures: > 451308 StoreEntries > 48956 StoreEntries with MemObjects > 48927 Hot Object Cache Items > 451254 on-disk objects > > I guess it is not a mem_pools_limit issue, since memory_pools_limit is > the default 5MB, which means the malloc library will handle memory > above 5MB. Not sure whether the malloc library returns freed memory to > OS or not (guess not). > > So the problem might be a rush of requests bring the host to its > knees, and it won't have a chance to stand on its feet again before > the squid restarts? Lots of swap in/out and few memory for filesystem > cache when squid holds above 90 percent of system. The memory > starvation simply exacerbates system load. > > Any solution? > > Thanks, > kc > > 2010/10/19 Amos Jeffries <squid3@xxxxxxxxxxxxx>: >> On 19/10/10 22:19, Kaiwang Chen wrote: >>> >>> Hello, >>> >>> In a 3.1.6 squid installation on a CentOS 5.5 x86 host with customized >>> kernel, top reports >>> >>> VERT 19.8g, and RES 14g >>> >>> while mgr:mem reports >>> Total Allocated 7.25g (96.4% of which is mem_node) high 7.27g, In Use >>> 7.25g %alloc 99.957 >>> >>> In squid.conf, the following is explicitly configured(mem_pool is >>> default): >>> cache_mem 7168 MB >>> maximum_object_size 51200 KB >>> maximum_object_size_in_memory 1024 KB >>> >>> Probably memory leak. However, it is not reliably reproducible. >>> Anyway, a restart will return memory to OS. >> >> see question 4 of the memory FAQ: >> http://wiki.squid-cache.org/SquidFaq/SquidMemory >> >>> >>> Additionally, mgr reports negative values: >>> >> >> Bit of 32-bit signed wrap in the display. Annoying but nothing to worry >> overly much about. It's been fixed in later releases. >> >> Amos >> -- >> Please be using >> Current Stable Squid 2.7.STABLE9 or 3.1.8 >> Beta testers wanted for 3.2.0.2 >> >
