[SOLVED] Disk IOPS vs Proxmox read ops

[Maher Khalil]

Hello
What is the difference between disk IOPS and proxmox VM read or write OPS/s
for example, is Samsung disk random read = 100K IOPS = 100,000 IOPS.
so, Proxmox read OPS = 100K OPS = 100,000 OPS

Is that true or I am in misunderstand?

 

[LnxBil]

In short: No, this is not the case.

Is that true or I am in misunderstand?

long version:
Those vendor IOPS are often block storage IOPS and do not relate to any "in-OS-IOPS" (in any OS). Also, those IOPS are "optimal" values and do often not relate to any real world usage, especially they are only valid for bursts (so short amounts of work). In addition to that, without knowning what kins of IOPS (especially the blocksize), the information is useless.

Futhermore, the storage architecture that you put on your block level SSD will reduce your IOPS further. Any filesystem will require some management overhead. If you then put a filesystem in a virtual disk on another filesystem, you will reduce the performance again. The number of simultaneous access will then further reduce (or enhance depeding on the used technology) the overall throughput. You will also have reduced read performance, if you also have to write at the same time (e.g. having access times stored in the filesystem). Another problem is the write and read amplification you have when updating one block on a cheap SSD, so that you will update a 4K block and have to read a whole 1 MB block and store that 1 MB block again. For ZFS as an example, you will have often significantly reduced performance, because there is a lot of logic overhead involved that gives your atomic snapshots and checksums for example, so that may not be the best high-performance filesystem there is, if you only want to have raw performance.

So you see, there is a lot of stuff to consider and in the end, you will have to do your own benchmarks with the SAME tool on different levels in your storage stack. Benchmarking the underlying block device yields normally the best raw performance, but is useless in real world application, because most applications need a filesystem (there are exceptions like Oracle ASM), so you need to benchmark on the filesystem level, then on the VM level and every filesystem you want to use has its own potential problems with respect to performance (e.g. parallel performance).

 

[Maher Khalil]

Hello LnxBill
I know that vendor number is not real, I do not want to discuss that, I want to know the relation between proxmox terms related to vendor terms.
I also aware about files systems issues
what I want to understand is vendor IOPS is the same proxmox OPS?

 

[LnxBil]

what I want to understand is vendor IOPS is the same proxmox OPS?

If the blocksize matches and you test with the same program under the same conditions, a random read IOPS is a random read IOPS. Yet you cannot know how the vendor benchmarks its own disk. Therefore you need to check that for yourself.

...and PVE is just Linux, so there is no PVE-specific IOPS.

 

[Maher Khalil]

I do not speak about numbers or matching between numbers. I speak about terms (word IOPS AT VENDORE IS THE SAME ops AT PROXMOX)

 

[LnxBil]

I do not speak about numbers or matching between numbers. I speak about terms (word IOPS AT VENDORE IS THE SAME ops AT PROXMOX)

If you only want to know if the shallow word IOPS (an input/output operation per second) in computer science is used in the same way as a dimension, then yes. But that's like comparing oranges and apples from orbit as I tried to explain to you, or better degrees of temperature ... if you don't know if they're fahrenheit or centigrade, you don't get any information.

 

[Maher Khalil]

Thank you