Is Azure Right For My Project?

Question

Is Azure Right For My Project?

R H 41

Hello: Am just getting my feet wet with Azure (the initial free subscription) and had some questions as to its capability. I wrote a program in vb.net that has to run through a routine millions of times. I have written it with 4 threads and created an Azure vm (Standard F4s v2) with 4 cores to run it on. On this vm the program is running about 1000 iterations per hour. Thus for just one million iterations that is 1000 hours. Running it on 2 such vms would be 2000 per hour or 500 hours. To get that down to a reasonable time would take many such vms. For instance, to get it down to even 10 hours you would need 100 such vms. Multiple 4 core vms does not seem like the answer to me. Question is whether Azure has any other type of solution that would run something like this in a reasonable amount of time. Would it work for me to run this on multiple 32/64 core vms? Note that this is a Windows program written in vb.net, which is what I know, however if there may be a solution that would work for me requiring a re-write in some other language or O/S, I would learn it. I really need some sort of super-computer type computing power (note that this is just computational intensive. there is little I/O or storage). If Azure isn't it, I would love to hear any opinions as to any other possible ways I might do this. Thank you. I appreciate your interest and help.

2 answers

Your answer

Answer 1

Michael Taylor 60,326

How about you remove the limit of 4 threads and let it use whatever is available on the machine? Then see if it is still slow. In general you should not code for a specific thread count as you have no idea what capabilities the processor has. For example your 4 core machine may support 8 logical threads. Furthermore your work may be a mix of IO and CPU bound work so it is possible while one thread is doing IO another thread can be using one of the CPUs. It is generally best to let the runtime figure out the threading and you just write your code to do the actual work.

You didn't post how you're breaking up your problem so we'll struggle to provide any guidance but unless you have a very specific need to control what work is done on what threads you might consider using the Dataflow library which allows you to partition large sets of data into smaller sets and work on them in parallel. It completely depends upon the problem you're solving of course but it is a good start.

R H 41 Reputation points

2022-11-23T04:57:30.827+00:00

Thank you for your response:

I have never coded for cores/threads or had any knowledge thereof. I thought "4 core vm. Need to code 4 threads" No idea you could have more threads than cores! Funny. Another member of this community told me I had to code for threads to get any speed increase. I ran it both ways. Without thread coding: 1000 iterations. 2.3 hrs. With thread coding for 4 threads: 1hr.

The program is practically all computational. Very little I/O or storage required. Loads about 60 meg of data into memory and runs a routine x number of times on the data based on a combination of parameters. It is the number of different parameter combinations that determines the number of iterations. It runs the same routine on the same data for each combination. The result is a small text file of results. The way I am breaking up the problem for threads is by breaking up the parameter combinations. For instance, for 1000 combinations, thread1 runs combinations 1 to 250, thread2 runs 251 to 500, etc. Each iteration runs against the full dataset.

Problem is the number of parameter combinations is actually in the millions, so neither of the above scenarios seem practical for getting results in a reasonable amount of time. The things I wondered were:
a) would it be doable/practical in Azure to run a number of 32/64 core vms with the program coded for the corresponding number of threads.
b) If not, does anyone know a method by which I, a retired programmer sitting at his home PC, can tap into supercomputer-level power (at an affordable cost).
Michael Taylor 60,326 Reputation points

2022-11-23T15:47:59.903+00:00

I think for your specific problem just buying a standalone machine and OS license (so you don't bring down your main machine) is the more cost effective option. Yes Azure can do it but you should look at the price calculator to see the cost. Maybe the first year it might be cost effective but VMs cost money even when they are off so you'll be paying that cost ongoing. As a quick example, an F16 VM (16 core, 32 GB memory) costs $1.732 / hour so pay as you go would be $727 a year. The OS license is $537 bringing the total per year to $1264. This doesn't include other costs like additional or faster storage, network transfers, etc.

Honestly I think your problem can be solved using a single machine with even 8 cores. You said your process takes 2.5 hours (150 mins) to run on a single core. Ignoring other processes, any cross dependencies and other things that impact multi-threading you could potentially get this down to 20 mins by using 8 cores.

Before buying hardware or using cloud consider using the Parallel class and see if it speeds up the work. If you need partitioning then use Partitioner to break up the work into blocks you control. If that still doesn't work then a dedicated machine may be in order.

Answer 2

Hello, @R H !

I'd like to second what @Michael Taylor has said--usually if you find yourself looking for supercomputer-level power for a personal project it's a good idea to look at optimization before looking at additional power. Azure does offer this kind of computing power but there's likely a better (and cheaper) way to go about this than by using brute force. I'd like to review some of the services that Azure offers which may be of use.

Is Azure a good option for my compute intensive workload?
There are very few scenarios where the cloud is not a better choice than an on-premises solution which is why so many companies have been moving to the cloud. A direct follow up question, however, would be what is the best architecture for your solution and to that end I'm going to include some Azure services that may be of use.

Advantages of Azure over an on-premises solution
For your solution, I imagine that cost and hardware availability are core considerations but there are a lot of other benefits that the cloud offers as well.

On demand: Solutions can be created on demand and deleted just as easily.
Setup time: Creating resources can usually be done within a couple minutes so you can get started quickly.
Cost: Pay as you go pricing with no upfront costs and competitive prices that often beat the cost of on-premises solutions (including things like electricity and cooling).
Scaling and flexibility: You can scale your workloads based on metrics with a few clicks (even automatically).
Power usage (and environmental impact): Heavy computation workloads use a lot of power so the cost of your power usage as well as your environmental impact will be important.
Maintenance: You can have Azure take care of your software and security updates leaving you more time to focus on your solution.
Reliability: Azure offers SLAs to ensure that your solution has as much uptime as possible.
Security: Cloud workloads suffer far fewer security incidents than those in traditional data centers.
Remote access and team controls: You can access your solution from practically anywhere, and on several different devices. There are also a lot of tools at your disposal when it comes to working with teams.

Read more: Cloud storage vs. on-premises servers: 9 things to keep in mind

Azure services for compute related solutions
There are a variety of Azure services you can use for your compute related solutions. If VMs aren't cutting it, you may want to look into a couple of the other options. Given the number of hours you are looking at, I might recommend starting at the bottom of the list with the big compute architecture style documentation that covers Azure Batch.

Virtual machines: VMs are a great solution for a wide variety of projects and it sounds like you are already experimenting with compute optimized VMs using the F4s v2.
Virtual machine scale sets: You can create and deploy thousands of VMs in minutes depending on your needs.
Azure Functions: You mentioned that you would be open to other languages and Azure Functions allows you to abstract away the operating system to focus on the code. Several languages are supported, including C#, Java, JavaScript, PowerShell, Python, TypeScript, and Go/Rust.
Azure Batch: Used for managing large-scale high-performance computing (HPC) applications.
Azure Kubernetes Service (AKS) and Azure Container Instances (ACI): Although these don't sound like a natural fit for your scenario, they may be worth considering for future projects.

Read more: Big compute architecture style

I hope this helps!

Share via

Is Azure Right For My Project?

2 answers

Your answer