Product SiteDocumentation Site

2.4.2. Latency and Threads of Control

The PCP protocols expect PMDAs to return the current values for performance metrics when requested, and with short delay (low latency). For some target domains, access to the underlying instrumentation may be costly or involve unpredictable delays (for example, if the real performance data is stored on some remote host or network device). In these cases, it may be necessary to separate probing for new performance data from servicing PMCD requests.
An architecture that has been used successfully for several PMDAs is to create one or more child processes to obtain information while the main process communicates with PMCD.
At the simplest deployment of this arrangement, the two processes may execute without synchronization. Threads have also been used as a more portable multithreading mechanism; see the pthreads(7) man page.
By contrast, a complex deployment would be one in which the refreshing of the metric values must be atomic, and this may require double buffering of the data structures. It also requires coordination between parent and child processes.


Since certain data structures used by the PMDA library are not thread-aware, only one PMDA thread of control should call PMDA library functions - this would typically be the thread servicing requests from PMCD.
One caveat about this style of caching PMDA--in this (special) case it is better if the PMDA converts counts to rates based upon consecutive periodic sampling from the underlying instrumentation. By exporting precomputed rate metrics with instantaneous semantics, the PMDA prevents the PCP monitor tools from computing their own rates upon consecutive PMCD fetches (which are likely to return identical values from a caching PMDA). The finer points of metric semantics are discussed in Section, “Semantics”