Recent Readings

The cloud is great. Stop the hype. – This is an excellent article on what cloud computing is and isn’t and when the use of the cloud is the correct technical or architectural choice. I had a long post planned on the overloaded term “cloud computing” but OmniTI covers all the important points in this article. Like any new approach to infrastructure deployment that promises quick provisioning of services, people often forget that all of that infrastructure needs to be managed. There are a lot of good tools coming out to help with that management but none make it zero cost.

Dissecting Today’s Internet Traffic Spikes – With the above article on cloud computing and this article on the sudden nature of internet traffic spikes, I’m becoming an OmniTI fanboy.  Part of my job is to worry about designing and provisioning correctly for sudden changes in traffic patterns, and Theo is correct that you have to design for spikes, not react to spikes.

Kanban For Sysadmins – I’ve started doing Kanban at work for one of our Operations teams and have been really pleased with the results so far — so much so that we’re rolling it out for another team this week and hopefully the rest of the department over the next few weeks. We track our work in Request Tracker, but it is hard to know 1. what is being worked on right now, and 2. how much throughput a team has. Kanban lets us know both, and it also lets us avoid the entire topic of prioritization of future work. We only prioritize when we are ready to start doing new work. I’ll post a follow-up to this once we’re further along in our Kanban experiment.

Hello From A libc-free World! – Have you ever wondered what, exactly, your “Hello, world!” program does? Jessica at Ksplice dives into what happens when you build a super-simple C program (it’s more complicated than you think!).

Data-Intensive Text Processing With MapReduce – A freely available draft in PDF of an upcoming book on using MapReduce to process large text datasets. One of the cool things we’ve done at ITA is add tracking data to each and every request that passes throughout our reservation system, and we output this tracking information in each log entry in every component we’ve written. The structure of this tracking data is such that if you aggregate the logs from all of the components you can easily construct a graph of the request’s path through the reservation system (including the asynchronous calls). The problem now is searching all of that log data, and I’ve been curious about MapReduce as it applies to this sort of data mining.

Performance Testing An Airline Reservation System

Until a few weeks ago I ran the performance and capacity testing team for the airline reservation system ITA develops. The group is under the umbrella of operations, which may seem out of place to many software shops, where typically the performance testing team exists in QA (or doesn’t exist at all until needed). We work very closely with development and QA as needed (and often, development has a dedicate set of engineers on performance work), and after doing performance work for the past few years, I’m convinced the best people for the job are the people that are skilled in development and systems administration (these are the DevOps people everyone is talking about). We’ve developed a lot of processes and tools to do our job and I think other people might find these ideas as useful as we have.

Testing Tools

At ITA we had to build many of the performance tools we use in-house because performance tools that could speak the airline industry protocols used by many interfaces to a reservations system (MATIP, for example) don’t exist. We also have a set of custom XML interfaces as well as a large collection of other interfaces that we need to send traffic to, or read instrumentation from. Our initial load generation script not only generated this traffic but also took care of all the other functions required to run an experiment, but this monolithic script didn’t scale. We ended up breaking up that script into agents that can be distributed across many machines, with each agent performing a single function needed for a load test. The agents are run by a master scheduling script which co-ordinates agent start and stop. In this way we can be sure that instrumentation requests aren’t blocking the load generation tools from working, and we can also schedule periodic events, report status, and do the hundred other things required for a full-system load test.

We gather a lot of metrics during a test, and for every major performance test we automatically generate a dashboard to help us drill into the results, a subset of which looks like this:



We gather this data from the system via SNMP, munin, per-component instrumentation, and other monitoring tools. We’ve been very happy with munin in particular as you can quickly add support for gathering new data types from remote hosts by writing simple Perl scripts.

Continuous Automated Testing

In any large system I’ve worked on the hardest problems are the integration problems, and a complex multi-component system such as a reservation system has these in spades. When we started doing performance testing, most of the system components weren’t finished and the interfaces between components kept changing. Furthermore, airline schedules, inventory and availability change rapidly over time.

There are countless factors that play into the performance and scalability of a complex system, and there are many philosophies around testing such systems, but in this post I want to discuss the technique that saves us the most time and money: continuous automated performance testing.

As discussed in the groundbreaking article Continuous Integration & Deployment In The Airline Industry [note: article not groundbreaking], ITA uses Hudson to build and test a complete reservation system on each check-in to the source tree (provided a build is not in progress). Hudson deploys the built software to a cluster of machines that are dedicated to continuous performance testing. After deployment, the load test master control software I discussed earlier runs a fixed scenario of load against the newly-deployed software. After a run completes, we store all of the results and instrumentation data in a database and update the graphs which trend test results over time. If our scripts find too much deviation in run time or throughput between this run and the previous runs, we set a status code so that Hudson can tell the people who’ve checked in since the last run that they may have broken the build.

Having a visual representation of performance issues in the continuous test environment has helped us tremendously because it both shortens the debug time and lets us see patterns of performance over time. Here’s an example of our throughput graph for a single component when someone breaks the build (click on the image for a larger version):

Along the X axis are revision numbers, and on our system the graph will show you the commit messages and the usernames of everyone who committed for each revision when you mouse over the data points.  We also make the graph very user-friendly with a “green lines are good, red lines are bad” design. Clicking on a data point will bring you to our internal source code repository browser.

Throughput, which is shown in the above graph, is only one side of the story. What about the run time of the system during the issue with revision 346626?

The multiple trend lines in this graph represent the timings reported by each instrumentation layer in this component. In the case above the graph is saying that the issue is not with CPU time consumed by the component (that trend is flat), but is instead with time spent in the database. This helps us quickly narrow down where to start looking for the cause of the performance problem. In this example, the developer fixed the issue quickly because the developer had notification of the failed test within an hour of check-in and had all the tools and data needed to isolate and resolve the problem.

At ITA we have environments we use to run large-scale performance tests, but the setup, execution and analysis for such tests are very expensive in terms of computers (many hundreds) and people (tens for what may be a few weeks for a single test). Those resources aren’t cheap, and the wins from automating performance testing finding a single bug save us more then the cost of the computers and people we invested in building this system — and we routinely see 2-3 performance regressions in a month.

It doesn’t take many computing resources to build a system like the one I’ve described. Here are some tips for doing this yourself:

  • Use real machines, as virtual machines suffer from the other guests on the same machine
  • Define a fixed workload you can replay via your load generation tool as this lets you establish a baseline to trend and alert from
  • Make sure your workload represents the majority of the types of load you’d see in production
  • Start simple and add metrics and instrumentation as you need them, not before
  • Don’t worry about fancy presentation of the results – it is more important that you start getting results
  • Publicize your testing system widely once it is up and running to help spread a philosophy of continuous testing in your organization

If you’ve got any questions I’d be happy to answer them in the comments and would love to hear about any systems like this that other people have built.

Recent Readings

Some interesting articles & tools: