IOPS, Throughput, Bandwidth Relation

[jam_1]

Hi Gurus,

Can anybody help me to find the definitions of IOPS, throughput and Bandwidth(2GB/1GB) and their relationship in a SAN environment.
What does it mean by a SAN device Spec of 5000IOPS,200MB/s, 2Gb.

Thanks
Bonny

[David Child_1 · Accepted solution]

IOPS = IO's Per Second
Throughput = Amount of data it can handle
Bandwidth = Network speed of the device

So the SAN device you listed can handle:

- 5000 IO's per second max
- 200 Mb per second max
- 2Gb per second max (per path I believe)

If you have an OLTP database it will generally have a fairly small block size (example 4k). To get 200Mb per second on this it would take a 51200 IOPS system (200Mb / 4k). This system only handles 5000 IOPS so the bottleneck in this environment would be IOPS.

If you are storing xray imaging files or something like that these would usually run at much larger block sizes (example 128k). Therefore you would hit the 200Mb/s limit before you hit the IOPS limit (200Mb / 128k = 1600 IOPS).

I believe the 2Gb item refers to SAN bandwidth. Its the size of the pipe to the array itself and could be a limit depending on # of hosts connected to it, etc. (e.g. all hosts need to connect to this system via this pipe (and share the bandwidth). So if you had 12 hosts trying to push 200Mb/s you would exceed the bandwidth of this pipe (12 * 200Mb = 2400Mb (which is >2Gb)). Of course there can be more than 1 connection into most systems. Note: If your switch only supports 1Gb then this unix would need to run at 1Gb speed.

Hopefully I got all the math and stuff correct. So basically when you analzye what you are looking for you need to take into consideration what will be running on the unit.

David

[Hein van den Heuvel]

Nice explananation. Agreed. Mostly.
I suspect the 200MB/s and 2Gb/sec refer to the same thing:

200 Mega Byte / sec and 2 Giga Bit /second

The divides out to 10 bits/byte.
A data byte is of course just 8 bits, but if you add in protocol overhead and stuff then you get close.
'stuff' is for example that a long series of zeros is NOT represented by 'lights out'. That would look too much like failure. So even for all-zero data, the protocol throws in a few ones be able to know where it is.

The 200MB could be the controller throughput, where the 2GB is the connectivity rate, with maybe muliple fibre available.. but the total still restircted to 200MB/sec.

2GB is where the current technology is.
If that is not enough, get more pipes in parallel.

5000 IO/sec is just 'ho hum' IMHO.
It will support just 50 disks comfortably.

Not too shabby, but there are controllers that can do 50,000 IO/sec.


fwiw,
Hein.

[jam_1]

Thanks David and Hein,
Yes I do think the same way as Hein about bandwidth.
200MB(throughput)*10(Number of bits)= 2Gb or it can be put as

Total bandwidth used by a host(for small BS) = Througput acheived*10= IOPS * Block size* 10

Thanks
Bonny

[Uwe Zessin]

2 GigaBits/sec vs. 200 MegaBytes/sec - 10/8 bits is not caused by the protocoll overhead or the difference between bandwidth and throughput. The reason is that every 8 DATA bits are encoded in 10 TRANSMISSION bits on the cable. It is for self-clocking and error detection on a serial media.

The protocoll overhead is 6.25% - so on the cable you would see 2.125 GigaBits/sec or 212.5 MegaBytes/sec.

[Rocky_8]

Agree with uwe,

And if im not mistaken this happens when the signal goes from parallel electric to optic serial --> The GBIC, where the 8/10 b encoding scheme is used.
Ive had quite the argument with "network" people on this x8/x10 calculation...

Furthermore, the way i understand it, its nice to have a device that handles a lot IOPS, like the EVA (168K p/s) but if the application is not servicing this or the block size is to small, good performance is not very likely.
The > the block size : the higher the performance as it takes less time to get more data to the device(s)= higher throughput.

Hope my vision of things helps somewhat

Rgds,

Rocky