I/OPS per VMAX 10K Engine

Yes I'm sorry the max possible IO rate.

Ok, pretty useless IMHO,

But I also need to know if it is a 10K (987) or a 10K (959)

Ok, but the maximum IO rate with 256K IO sizes with a 50/50 read/write ratio will be very different from a 512 byte read hit rate, maybe by 10 times or more.

Also still need to know if it is the newer 10K platform or not.

987 or 959 should be the first 3 digits of the serial number.

The CPUs are faster on the 987.

Yes, the IOPs per second goes down with larger IOs, but the MBs/sec go up.  So it depends on if you are trying to deal with small OLTP type transactions with small block sizes, or data warehouse, with large block transfers of a lot of data.

I'm sorry I'm not really sure how to find out what you're asking. We are slated to go to training very soon, I've been learning on the Unisphere for the time being. I've gone into the system dashboard and viewed the details but it doesn't give me a very specific model number. I can tell you it was just installed a month or so ago. (Image attached is what the "View Details" option displays)

Also I'm trying to make sense of your graph. It would seem to imply that as your read and write sizes grow your over all IO goes down? This would seem counter intuitive to me, I always thought that the larger your read and write blocks got the more efficient it was to write them to disk.

Thank you so much for all your assistance, these forums really are wonderful especially for folks like me who are just starting out.

Steven

An example of IO size vs. IO rate

So let me give you a better understanding of why I'm asking the question because yes I agree this is totally useless. We are looking at on boarding some data, about 50/50 read write ratio, and more reads and writes are about 256k in size. It's a very specific medical application with consistent data.

The potential client is saying right now they are spiking up to about 10,000 I/OPS on their current storage array. I'm trying to understand from an engine perspective what we would need. I'm just starting with raw numbers.

Thank you much,

Steven

Thank you, we're 987. We are dealing with an OLTP situation, so that would seem that we would get more I/OPS due to the smaller transaction sizes.

Quincy thank you so much for all your knowledge!

As far as i can recollect, you mentioned in your presentation (and the rule of thumb that we follow) that anything above 256KB should be avoided ( or try to avoid ).

So my understanding is the max IO value will be a relative value depending on the workload characteristics rather than an absolute number.

Ok, earlier you mentioned the IO size will be about 256K.

So the maximum IO rate with 512 byte blocks, read hits, for a single engine VMAX 10K (987) will be around 200,000 IOPs or about 25,000 per front end CPU.

If the IO size is 256K, the maximum rate for read hits will be about 40,000 IOPs or about 5,000 per FE CPU.

If the R/W ratio is about 50/50 and the IO is read miss and write new slots, the maximum IO rate will be around 11,000 IOPs for the system, or about 1,300 per FE CPU.  This may not sound like a lot, but that is 2.6GB/sec.

The throughput gains are minimal, if any, using IO sizes larger than 256K.  Typically the IO size is increased to gain extra MB/sec or GB/sec. As you can see in the chart, we do get better data throughput with larger IOs, but there is a point in which there is diminishing returns.

Hi Quincy56, considering the new VMAX 10K uses 2.8 GHz Xeon Westermere, can you please demystify the computations.

Thanks for the clarification.

These were not computations, but based on actual test results.

There can be two basic bottlenecks for performance in the Symmetrix, the CPUs or the plumbing.  Usually with very large data transfers, the CPUs are not 100% utilized, but the paths or plumbing that connects them are.

As for the front end CPU performance, this depends on which of the 4 IO types it is performing.  Each one takes a different amount of commands to complete, or CPU cycles.