SystemMemoryUsageLoadMetric is not correct on Linux/Unix -------------------------------------------------------- Key: MODCLUSTER-288 URL: https://issues.jboss.org/browse/MODCLUSTER-288 Project: mod_cluster Issue Type: Bug Affects Versions: 1.0.10, 1.1.0.Final, 1.1.1.Final, 1.2.0.Final, 1.3.0.Alpha1 Environment: Linux/Unix operating systems with standard memory management Reporter: Michal Babacek Assignee: Radoslav Husar Fix For: 1.3.0.Alpha2 I am convinced that the [SystemMemoryUsageLoadMetric.java|https://github.com/modcluster/mod_cluste...] is of no use to anyone. I have tested it by allocating some memory outside the JVM control with [NativeMemoryStress.c|https://github.com/Karm/clusterbench/blob/system-mem...]. The results from Mod_cluster are such that the system is almost fully loaded all the time despite the fact that I work with 8GB RAM box with almost nothing running on it. This [picture|http://hudson.qa.jboss.com/hudson/view/mod_cluster-QE/view/mod_cl...] displays time on X axis and (100-(Load from Mod_cluster)) on Y axis. As you may observe, all 4 nodes appear to be under full load all the time and there is a spike on each of them. This spike represents a period of time when the aforementioned [NativeMemoryStress.c|https://github.com/Karm/clusterbench/blob/system-mem...] was called with the following setting: {noformat} megabytes 500; milliseconds 60000; {noformat} What had happened is that kernel has cleared some cache and garbage data from the RAM in order to accommodate new demands, so by the time I freed the memory, there was all of sudden actually more free RAM then before. The actual problem with the current [SystemMemoryUsageLoadMetric.java|https://github.com/modcluster/mod_cluste...] implementation using [OperatingSystemMXBean|http://docs.oracle.com/javase/6/docs/jre/api/manage...] is that this code: {code:lang=java} +++ public static final String FREE_MEMORY = "FreePhysicalMemorySize"; public static final String TOTAL_MEMORY = "TotalPhysicalMemorySize"; +++ double free = ((Number) this.server.getAttribute(name, FREE_MEMORY)).doubleValue(); double total = ((Number) this.server.getAttribute(name, TOTAL_MEMORY)).doubleValue(); return (total - free) / total; +++ {code} Is somewhat similar to this (on 8GB RAM machine with almost nothing running on at the moment): {noformat} [karm@killerbox karm]$ free -m total used free shared buffers cached Mem: 7933 6755 1177 0 44 141 -/+ buffers/cache: 6569 1363 Swap: 16383 58 16325 {noformat} Obviously, this does not mean that I could get only 1177MB of RAM if I want to. The "free" actually means "wasted". The majority of RAM is being used as cache. If I ask for e.g. 2GB, I am gonna get them, because kernel will simply sacrifice some cache, clears some garbage from RAM and free up space so as to accommodate my demands. So, according to my opinion, on a sane, well-behaving caching operating system, the [SystemMemoryUsageLoadMetric.java|https://github.com/modcluster/mod_cluste...] will always show ~ 98% RAM usage. Due to this fact, this metric is unlikely to be of some use to anyone on Linux/UNIX. What do you think?