ORIGINALLY POSTED 31st October 2007
12,098 views on developerworks
I promised a bit more detail on the very limited sample set used by EMC’s latest blogger Dr Kartik, in response to his analysis that SPC-1 IOPs is directly proportional to the number of drives used by the system under test.
So I cannot deny the fact that :
larger number of drives = larger IOPs result
It is quite clear, and pretty obvious that the limiting factor in most of todays storage hardware offerings is the number of drives it can support. Each drive will have a fixed number of IOPs based on its rotational speed and interface characteristics. This goes back to one of my first blog posts. However as OSG has tried to explain, and Dr Kartik confesses he expected, SPC-1 would show at what point either the hardware or the disks become the limit. DrK’s comments on BarryB’s post and in his original post suggest that its only measuring the disks. I would say that its quite apparent when you look at ALL the results, not just a few carefully chosen numbers, that it is measuring both aspects.
I have plotted all the results available on the SPC website for SPC-1 and also drawn the trend line which quite clearly proves the assumption above and the main gist of DrK’s post. HOWEVER… (click on image for full size pic)
[Unfortunately, developerworks has lost the images from this far back]
I’d also speculate that all this proves is that the top performers in the benchmark all can make approximately the same utilisation of the disks at their disposal. SVC is a kind of outside in this comparison. Here the benchmark really is proving what the limits are of the SVC itself. The storage used in our configuration is mid-range based, and our EMC friends are quick to tell us that SVC is not an enterprise product, and so by the ‘trend-lines’ set out in DrK’s picture should really be included with the second trend… Its also no co-incidence that we had to use more drives, we carefully calculated what the limits of SVC would be and assembled enough disks to ensure this was actually possible. We have the luxury of adding as many disks as we need, whereas controllers themselves have physical limits. I would therefore expect to see a perfect corelation between the number of drives used to SVC IOPs figures as we will need more and more drives behind the box as we continue to enhance and utilise the latest and greatest developments in the xSeries hardware platform over the years. So while SVC is always compared against products like USP-V and DS8000 it really is a very different beast. We would probably even be compared against DMX, if EMC backed up their speculation by submitting. This puzzles me. Its not like EMC to miss a trick and miss out on a potential bit of ammo to sling around in their sales discussions – if DMX really was capable of driving its 2400 DDMs to almost 400K SPC-1 IOPs… but as I’ve been told many times, its irrelevant, and its just not going to happen. (Their words, not mine).
Anyway I digress. So back to the numbers. There are several points on these two graphs that show its not just disks alone that will limit, or enhance performance. Look at the DS8300 vs DS8300Turbo numbers (1). Here they both use 512 disks, yet the Turbo model improves on the previous result by some 25%. That would surely say that the box itself does have a major impact on the numbers. Now, I didn’t set out to show anyone in bad light, just prove that SPC-1 is more than a function of the number of drives used. Look at the 3PAR box for example (2), over 900 disks yet WAY off the trend line. This would say that this box is limiting the performance well before the drives reach their limit.
same can be seen at the lower end of the scale. Here there is a much
wider variation in the number of drives vs IOPs. There is still the same
general trend, however each box fluctuates to greater proportional
degree, both above and below the trend line. As an example, the Fujitsu
Eternus3000 m600 (3a) vs Eternus3000 m700 vs IBM DS4800 (3b), all using
around 230 disks. Same goes for the IBM DS4300 (4a) vs IBM DS4700 (4b)
using the same number of drives. I think I’ve made my point.|
Anyway, welcome to the blogsphere DrK and I look forward to future discussions, I guess in a way I agree that the basic principle of more drives = more performance, but SPC-1 can tell you more than just that.
As a foot-note, I was directed to the paper by Randy Kerns from as far back as 2005 where he concluded the same basic assumption but explains in some detail, with the data to back-up the claims, why this actually gives the ‘efficiency’ of a system. The paper is available after you register with www.evaluatorgroup.com and is titled “WP0024 A deeper look at storage performance”
All results plotted here are taken from the ‘executive summary’ reports freely available from the Storage Performance Council’s website. Please see http://www.storageperformance.org for more details and all benchmarks results in detail.
For those products cited above, please see:
IBM DS8300 : http://www.storageperformance.org/results/benchmark_results_spc1#a00044
IBM DS8300Turbo : http://www.storageperformance.org/results/benchmark_results_spc1#a00049
3PAR S800 http://www.storageperformance.org/results/benchmark_results_spc1#a00027
Fujitsu Eternus 3000 m600 :http://www.storageperformance.org/results/benchmark_results_spc1#a00014
TFujitsu Eternus 3000 m700 : http://www.storageperformance.org/results/benchmark_results_spc1#a00032
IBM DS4300 : http://www.storageperformance.org/results/benchmark_results_spc1#a00017
IBM DS4700 : http://www.storageperformance.org/results/benchmark_results_spc1#a00046
IBM DS4800 : http://www.storageperformance.org/results/benchmark_results_spc1#a00050