Dear Cephalopodians, for our MDS data, we are currently using 4 Bluestore-OSDs with classic SATA SSDs (Replica mode). In metadata stress testing, this easily becomes the bottleneck. Also, the existing SSDs are too small (4 x 240 GB, 4 replicas) if the cluster was to become full. My question is: Is it more reasonable to buy small, but fast NVMe devices for a WAL, and larger (cheap) SSDs for the KeyValue-store, for example 8 SSDs to also have more OSDs to share the load, or would this not improve the I/O situation, and we should really buy large and pricey NVMes? (The main issue with that is that our current servers only have a single free PCIe slot for that, so scaling that is expensive.) The answer strongly depends on how much I/O actually goes to WAL and how much goes to the DB. Is this the same I/O by design, or does the DB device need less actual IOPS (i.e. only throughput) due to the deferred writing? Anybody with experiences on this? Cheers and many thanks, Oliver
Attachment:
smime.p7s
Description: S/MIME Cryptographic Signature
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
