Windows Azure Scaling for I / O Performance

Question

Windows Azure advertises three types of I / O performance levels:

• Extra Small: Low
• Small: moderate
• Medium and High: High

So, if I have an IO-bound application (not a processor or memory), and I need at least 6 processors to handle my workload - will I improve I / O performance with 12-15 Extra Smalls, 6 Smalls or 3 mediums ?

I'm sure this is application dependent - is there an easy way to test this? Are there any numbers that give a better idea of ​​how much I / O performance gain you get when switching to a large instance role?

It seems that the I / O performance for smaller roles may be equivalent to larger ones, it is only those that are first reset if the total load becomes too large. Does this sound right?

Answer 1

Windows Azure Compute Size is approx. 100 Mbps per core. Extra Small specimens are much lower at a speed of 5 Mbps. See this blog post for more details. If you are attached to IO, the 6-Small setup will offer much more bandwidth than 12 Extra-Smalls.

When you talk about handling your workload, do you work in line? If so, then several work roles, each of which is an instance of Small, can then work with the 100 Mbps protocol. You will need to benchmark to determine if 3 Mediums provide sufficient performance gains to justify a larger VM, knowing that when your workload is reduced, your “unoccupied” cost per hour is now 2 cores (average, $ 0.24) USA)) vs 1 (small, $ 0.12).

Answer 2

As I understand it, the amount of IO allowed for the kernel is constant and needs to be allocated. But I could not get official confirmation of this. This is probably different from x-small instances that work in general mode and do not stand out like other instances of Windows Azure vm.

Answer 3

I would suggest that you actually think that even IO binding is application dependent. I think you could achieve your synchronization goal by using timers and writing output to a file on storage that you could retrieve. Define the math to understand that you can handle X the number of work units / hour, sorting as much through a small and then an average instance as much as possible. If the size of your work unit varies greatly, you may need to average too. I would always prefer smaller copies, if possible, and just increase the number of copies as you need more firepower.