The following two queries can be used to graph the average latency of the cluster. Larger I/O sizes per operation will increase the average latency, as larger I/Os take longer to process. As such, these graphs will not give a clear picture of your clusters latency if average I/O size changes over time. We have the following code:

    averageSeries(collectd.osd*.ceph.osd.*.ceph_latency.Osd.opWLatency)

    averageSeries(collectd.osd*.ceph.osd.*.ceph_latency.Osd.opRLatency)