> > ----- Le 14 Mar 25, à 8:40, joachim kraftmayer joachim.kraftmayer@xxxxxxxxx a écrit : > >> Hi Thomas & Anthony, >> >> Anthony provided great recommendations. Danke! >> ssd read performance: >> I find the total number of pg per ssd osd too low; it can be twice as high. They’re short on RAM, though, so I hesitate to suggest that. YMMV. >> >> hdd read performance >> What makes me a little suspicious is that the maximum throughput of about >> 120 MB/s is exactly the maximum of a 1 Gbit/s connection. Modulo framing bits and collisions, but good observation. I’ve seen a K8s deployment where some workers anomalously had a /32 PREFIX in interface configurations that sent intra-subnet traffic up to the router for no good reason. >> (I have seen this in the past if the routing is not correct and if you use >> VMs for testing the network could be limited.) Now that I’m not typing on my phone … this informs checking the networking on all nodes: Assuming that your links are bonded for only a public network: * inspect /proc/net/bonding/bond0 on every host, maybe ethtool on each physical interface as well. * Are both links active? * Are their link speeds as expected? * Use iftop or other tool to see if both links are getting very roughly comparable traffic. * Does netstat -i show a high number of errors? _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |