Cloud-based applications are often performance sensitive set of services deployed in virtual machines(VMs) that are provisioned either having all tiers together in a single node or spread across multiple node within the data center. Since cloud providers at the IaaS level see these VMs as mere application black boxes and there is hardly a means to identify the dependency between them, it becomes very difficult to first identify inter-vm dependencies and secondly to identify which of the tier is the source of any perceived performance drag.
I don't think that it is feasible to detect those (as you call it) dependencies entirely automatically, but maybe a set of smart heuristics along with some domain knowledge (e.g., standard ports of various services) and monitoring of network traffic could allow you to at least mine some potential cloud topologies (which you would, of course, show to the service provider / operator for confirmation or detail changes).
Detecting the source of performance problems is a somewhat unrelated problem, in my mind. I would start with simply looking at the standard performance metrics (CPU utilization, etc.) of the hosts running the various components. But seeing that you are at Umea University (working with Erik Elmroth?) you probably have a lot of knowledge in this domain in your group anyway, right?
Thanks Philipp.
First, yes I work with Erik Elmroth.
Secondly, about bottleneck, I quiet agree with you that a totalitarian solution to the dependency detection problem is almost implausible, I however think a near optimal solution could be achieved using analysis of a combination of performance metrics. I have found a few existing works in that line but want to know who else is working in that area. Whichever way, this is one of the problems I am currently working on....as this will by a long shot improve the management of multi-tier apps in the IaaS cloud.
Thirdly, it is our believe that solving the first problem will help us address the bottleneck issue as identifying source of performance bottleneck in multi-tier apps is made even more difficult by the dynamics of the cloud itself. At the moment there is no trivial solution to it as you think. Anyways, we are interested in development of a technique based on statistical analysis, heuristics, and machine learning techniques for addressing this issue.
I thought some more on this question. After further thinking, I am wondering what problem this dependency detection solution will solve. I mean, is the assumption here that developers are not aware about the topology of their multi-tier application? Because, honestly, that would seem like a stretch.
Existing state of the art, say TOSCA, AWS CloudFormation, or OpenStack Heat, clearly assumes that the developer / operator herself is able to easily specify the application topology (which includes the dependencies between components, e.g., "web server database"), and so far I have not seen an indication that this is in fact a problem in practice. If you are not aware of these works, you could have a look at this book chapter: http://www.iaas.uni-stuttgart.de/RUS-data/INBOOK-2014-01%20TOSCA%20Portable%20Automated%20Deployment%20and%20Management%20of%20Cloud%20Applications.pdf This describes the TOSCA Cloud Topology language in some detail (it's an upcoming OASIS standard).
With the third paragraph I wholeheartedly agree. I am also very interested in performance measurement and prediction for cloud applications. Currently, I am mostly looking at benchmarking and ranking public cloud providers, but looking at performance prediction for complex multi-tier applications is somewhere on my research radar. You can keep me in the loop if you achieve any published results in that!
Thanks once again.
About the dependency issue, our assumption is on the side of the cloud provider as it is often the case that his customers may not be able to technically identify this dependencies and so the providers merely see applications and services running on the infrastructure as a pool of black boxes. Given such assumption, it is believed that a dynamic approach to the detection of this intelligence will enable providers make better informed decision regarding performance management(migration, consolidation, resource allocation etc) for the various applications. However, we have given it a thought that a hybrid approach (static and dynamic) would be most practicable. Our assumption is further strengthened by the fact that.....making static dependency definition for a few apps is easier but becomes laborious with hundreds and thousands of apps from the same source.
I appreciate the reference you sent me, I will give it a read, but it cut across to me as not directly targeting Virtual Machine based apps.
Sure I will keep tab on you as we progress.
Thanks again. Been wonderful reading your comments.
"I appreciate the reference you sent me, I will give it a read, but it cut across to me as not directly targeting Virtual Machine based apps. "
Oh yes it is. TOSCA specifically targets applications deployed in a complex architecture among multiple virtual machines. Anyway, I understand now that you are talking more about discovering these dependencies from the point of view of the cloud service provider, not from the point of view of the developer / operator (which makes more sense).