|
300 servers are stretching the practical limits of pmwebd's current approach to searching archives, especially if the archives are large enough not to fit into RAM. In fact if the active set of archives (those that pmwebd needs to read, and those that something else (pmmgr/pmlogger?) is writing) are too large to fit into RAM, then I/O will start dominating everything, as you are noticing. Can you offer some stats about your archives? How far back do they go? How large are the currently-written-to ones? How much RAM do you have? How many separate archive files exist? Are any of them compressed (via service-pmlogger's pmlogger_daily, as in *YYYYMMDD.0.xz)? see also — |