ceph performance estimation calculation?

[ilia987]

we are planning to add a new ceph pull that will be consists of 60 HDDS across 5 servers with 40GB duel network. (one for ceph sync, and one for clients)
in future all the hdd slots will be populated

how far this assumption from the reality:

• HDD have a read speed of 100MBs
• the data is accedes by multiple clients (around 30-50 ) every time. each need around 100-300MB of data
• the use case is read only
• replicate 3

can i assume that the data read will be 100*60 => 6GB sec?

 

[Alwin]

can i assume that the data read will be 100*60 => 6GB sec?

No, not all disks will be accessed simultaneously. And writes will interfere with reads. Also the latency will have an impact on the overall performance. Then the question is, how concurrent the reading is and what the pattern is with 50 clients. In the case of read-caching much performance can be gained and can not only be done on the Ceph level.

 

[ilia987]

there will only read requests. (no write while the data is used).

the 50 clients is just an estimation. it total we have around 500 clients. but each need a cunk of data once every 10-40 seconds..

caching is useless because he data that needed to be read is not repeated.

the data is used for simulation. and each client need a separate sequential chunk.

the current storage that we have is all flash. and the speed is more then enough. but we use fast sas3 ssds. but the cost is too high to match our scale needs.

 

[Alwin]

More nodes will less disks per node will push the performance. But in the end you will need to calculate the numbers with what performance you need to what hardware you want to buy.