Slashdot Mirror
Debugging Asynchronous Applications?
duncan bayne asks: "I'm attempting to debug a complicated telephony application, written in C#, that's almost entirely event driven. This is the first time I've debugged a large asynchronous application that isn't a GUI, and I'm curious to know what advice the Slashdot crowd has to share - have you any recommended tools, best practices, or common pitfalls to avoid?"
One other suggestion... "event bus" apps like you describe are good candidates for capturing as much runtime data as possible, so make sure you adjust your build parameters and do as much of that as possible, especially in problem assemblies. Oh, and don't forget to build nUnits. Sounds like you're walking into some prewritten code, but the effort might be worthwhile.
Sigs cause cancer.
As a basic sort of testing phase, do it all on one computer. This eliminates all possible network errors that can occur. I'm assuming this is meant to be huge so maybe the bugs you speak of result from multiple machines fouling each other up. Either way, let's talk debugging strategies!
... yet. Their papers may be of use to you, however. If you're having problems with packets on either end, use my good friend ethereal.
Also, as I recall from my days of drudgery at college, create tons of output.
So I will suggest as a preliminary requirement that you create a nice logging system (if you haven't done so already). I haven't written much C# so I'm going to be talking abstractly. Hopefully the rest of Slashdot can help with the specifics to C#. Now, what I mean is that you should create a class that just creates an output log file that you can read for output later. I don't mean to put a message for every packet sent but maybe it wouldn't hurt to put a message for each stream or connection opened. It's going to help for you to generate random IDs for each call and to put the destination/receiving IP:Port in your log. This would most likely be helpful with a server. It also will be helpful to store printlns in your code (redirect standard out to the logger).
Now use this on every machine in the system. If one machine should start to give you problems, create a mutual exclusion on this log (or put all of the log entries in critical regions). In Java, you can use object locks or the synchronized keyword--in C# I'm pretty sure they have something similar. Just because it's not a GUI doesn't mean you can't record output.
Just a friendly warning, time stamping is usually worthless unless you have a logical network (i.e. a Lamport Clock) clock scheme set up (which usually requires lots of time on one's hands). You could shoot for an NTP server but I wouldn't trust the accuracy past 500 ms. If you absolutely need a clock scheme, I recommend having one machine on the network tick tock an increasing number that is reflected in all the logs. Make the time between ticks adjustable--this way you'll be able to check out events roughly relevant to these ticks (assuming the time it takes to get there is similar).
In the end, your best tool is your brain. Designing tests and double checking the logs on each machine to see that the linear time sequence of relative events is correct. Logic will be your only friend in this journey. Don't be afraid to kick off more threads on the client side if they don't need to share resources. If you have a server side, be careful in how many threads you have and make sure you realize what memory scope they're limited to.
For the love of god, if you use ports--don't forget to free them when you're done using them!
Unfortunately, Nornir is not OSS
Good luck! Happy debugging!
My work here is dung.
The buffering nature of printf along with the asynchronous execution of code can lead to out of order debug printfs.
I had this one project. It was to build a model car, not related to programming at all. I started out doing well, following the instructions and generally getting along fine. But then I lost patience with the tedium and left to get a beer to relax. When I finally got back to working on the model, I found that the dog had chewed it up and the wife had thrown it out (the trashed model, not the dog, but she'd love to throw out the dog too). I left it in the garage where I thought it would have been safe, but I guess you can't expect things to stay the same if you leave it sitting there for a year and a half.
The moral of the story is that if I look to see where things went wrong, it was the point where I lost patience and decided to do something different than what I should have been focused on. This is like how many people try to put breakpoints all over their code rather than where they should put them. Don't debug willy-nilly and expect to make any good progress. But also don't try to throw in some seemingly helpful actions (like printf) because it may end up changing the whole state of the program.
Chapter 1.
Stand up and slowly back away from the keyboard.
Chapter 2.
There is no chapter 2.
A great reference for design in these types of applications is Enterprise Integration Patterns By Gregor Hohpe and Bobby Wolfe. Some of the paterns contained provide methods that you can use to debug and trace such appliactions. Typically, these apps deal with messages as the primary unit of transport. You have kind of a heisenberg uncertainty principle in effect, as you can tell where a given message is in the system, or what the flows of through a given point of the system, but it is hard to tell both. And the at of monitoring a system can often affect timing, and change behavior. One thing you can do with messages is put proxies in place that allow you to log every message delivered to a given componenet. Other tricks are to add routing slips to messages that indicate what componenets they have passed through, but this might not be possible to introduce into an existing system. Network tracking tools are invaluable, netstat and tcpdump are your friends. Or whatever tools work with your particular network stack.
Open Source Identity Management: FreeIPA.org
I've found and fixed more bugs in large systems by analysing logfiles than by running tests in a debugger. Log everything, and make sure each line in the logfiles has an accurate time stamp.
The trick is to learn how to correlate information between different logfiles to build up a picture of how all the components (process or thread) behave together. The classic Unix utilities like find, grep, awk, cut and less are your friends.
The question is kind of vague. Are you debugging receiving events? Sending events? Both? I'd look into Mock Object and see how they can help simulate asynchronous environments.
Look for invariants and ensure that they hold true where they are supposed to. That doesn't require fine analysis, but will detect problems in the logic when you're running the full system. Use profilers and coverage analysers to make sure that when you DO do invariant checks that you're actually checking all the areas they're supposed to hold up.
Test "normal" values, borderline/extreme values (that's where overflows, underruns and other assorted nasties are likely to show up), and completely erronious data. Borderline can include borderline values, but since this is a network app, it can also include extreme volumes (very little or vast amounts of traffic, or even massive variations). Erronious data can be data that is invalid in and of itself (malformed packets, for example), or data that has no rational meaning (a non-existant codec, or a value that decodes to something absurd - I doubt many people will be doing 11.1 audio streams for VoIP, for example!)
Beyond that, it's all much of a muchness. There's very little that is async specific to testing, so long as you concentrate on the logic rather than the means of getting there.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
I've done a bit of asynchronous debugging, principally troubleshooting interaction between a speech recognition system and Emacs. Most of the bugs I found tended to be due to errors in assumptions in the code. For example,
(a) when a "change text" event handler in the editor is invoked, the editor will always be done reporting the result of the previous change.
(b) event z will always be preceded by event w
If you know the assumptions for each event handler, cases where they break down may become obvious. If not, you can add assertions to check those assumptions.
Beyond that, strategies really depend on what sort of feedback loops you have. When you are debugging interaction between two objects or programs, try to simplify one so that you don't waste a lot of time trying to figure out which side is buggy. For example, I wrote a simple brain-dead editor which I was confident I understood and connected it to the speech recognition system. That way, when I found a problem, I was confident that I could quickly find any bugs on the editor side, so if I didn't find the bug there quickly, it had to be in the speech reco system.
I have mod points, but I don't see anyone chiming in here about realistic load testing.
For this kind of application, you must, *must*, MUST create a heavy load on a production system. I've done work with big, complex, multi-threaded web apps that have similar characteristics -- event-driven (when an HTTP request comes in) and server-only (no GUI). There are many bugs that don't show up until you put the system under load, as in dozens or hundreds of transactions per second. For instance, under light load a queue will never fill up, but under heavy load bizzarro, difficult-to-trace bugs will crop up that you can't reproduce on your development system. Even under the same load, your development system may run into a different constraint (e.g. CPU-bound so that it can't fill the queue fast enough and thus never hits the bug).
To have any hope of catching these bugs, you need to instrument your application heavily, with logging calls that you can turn on and off easily with some sort of switch (kill signal, special dialing code, etc.). Running with a debugger attached will likely be next to impossible on your production or staging systems.
Lastly, definitely invest in an automated test environment. You will need to do these kinds of debugging runs hundreds of times in the course of developing your app, and it just isn't feasible to have everyone in the company drop what they're doing and call into your app a dozen times a day. While there are plenty of load test tools for web apps, I'm not familiar with any for telephony apps, although some must exist. You may end up rolling your own from a bunch of old modems.
Good luck, as the bugs in these systems are notoriously difficult to hunt down.
--Paul
Asycn application? Do you meant you have two parts of a client-server relationship, such as an SMTP client and server, or a more direct (and less formal) communication channel like shared memory?
By event driven, do you mean that you have events requiring immediate attention, or do you have events you can buffer with water marks?
Is it a direct producer-consumer with only 1-way communication, or do you need bi-directional communication?
If you give some one a poorly worded problem, they cannot be expected to solve it.
--
Internet Explorer (n): Another bug -- that is, a feature that can't be turned off -- in Windows.
For what it's worth:
1 - be cautious about testing debug mode - there's an awful lot the compiler tosses in to enable debugging which may impact how the code actually executes.
2 - use logging extensively. I'd recommend using log4net or something like that.
3 - use an integration model for your unit testing. Start with the smallest unit tests and build upwards. This will allow you to gradually build "correct" code and focus on the messages/events between components.
4 - build a simulator (someone mentioned that before), they are truly invaluable. Keep it as simple as possible.
5 - check, double check, and triple check variable access. It's easy to run into a race condition between reads and writes. Study and understand lock(...), reader-writer-locks, semaphores, and mutexes.
6 - when testing, don't forget to test expected conditions, unexpected conditions, boundary conditions (null objects, empty strings, negative values, 0's, positive values, and overflows), errors (like zombie conditions where a response is _never_ generated, dropped connections, garbage results), etc.
7 - learn Debug.Assert and check your pre- and post-conditions
8 - if you use strings, make sure you understand how strings and stringbuilder operate - they can have dramatic differences in efficiency/memory utilization/and GC.
9 - events can be static, and don't forget to encapsulate your event accessors (they look like properties, but instead of get/set, they're add/remove)
10 - if you plan on using compiler optimization switches, use them last during testing - after you can prove the app works correctly. Optimization switches can dramatically reorder things which is definitely not good if you're trying to determine correctness.
11 - set the compiler to give the maximum warning level. Your app should generate no warnings or errors while compiling.
12 - walk through the code with someone to double check your logic and field access. If you can convey it to someone else either through comments, design, etc., and can justify all field accesses as well as access control, you'll be in good shape. Yes, this is peer review. It's even useful to haul in a project manager that knows nothing about coding and nods like a chicken. Listen to yourself as you talk and if you stumble or have a hard time explaining something, that's a hint that a redesign might be in order.
13 - you might want to put in some profiling counters so you can capture metrics on it. This way, as you change the code over time, you can almost quantifiably determine if the code is truly improving with respect to throughput, responses, etc. or not.
That's all I can think of off the top of my head.
Good luck, it's a fun journey.
[Of course it's client-server; it runs on a LAN]
I'm unfortunately not too familiar with C#, so I can't comment on it's logging facilities (or lack thereof) other than the .NET EventLog class.
There is a project on Sourceforge called C# Logger that is supposedly similar to log4j in Java. But it seems to be stuck in alpha release mode, and not particularly active.
Just my two cents. Hopefully it helps. :)
-- null
50 comments, and not one good answer (though I saw three posts of good advice vaguely applicable to your needs).
.NET to
gracefully handle network errors. Don't trust your process
to have basically-uninterrupted control of the CPU. Don't
trust try/catch to save you from "real" problems.
First of all - Debugging takes hard work. Sorry folks, no matter how easy Microsoft tries to make it, no matter how tightly they integrate Java-killer-P into app-Q, you still need the ability to follow the flow of bits from point A to Z, and more importantly, figure out what B through Y need to do.
How to debug asynchronous events... Since you mentioned c#, I will presume you have a REALLY coarse granularity here when you say "async".
So... First step, force non-reentrancy and non-overlapped event handling. Does your problem go away? Find the global data you clobbered.
Step 2 (if #1 fails) - Run both ends of your app on the same machine. Does the problem go away? Don't trust
Step 3 - Okay, you have a "real" bug, in your code. But on the bright side, if you got here, you can probably reproduce it, so, piece of cake. Load up your trusty debugger and dump a COMPLETE stack trace up to the error. Don't trust the last line to have caused the error, it just failed to deal with whatever broken crap the actual problem threw its way.
Step 4 - trace through your code, on both sides, one line at a time. Sound tedious? Yup. You might spend a week on a single run. But you'll sure as hell know the flow of your by the time it finishes.
Step 5 - No step 5 exists. Step 4 WILL let you find your problem, as long as it resides in your code and not in the aether between the two sides of the connection (which steps 1 and 2 should have eliminated as a problem).
. . . deadlocks are going to be your enemies. Especially deadlocks.
An easily vanquished enemy. Just log into Visual Haircut 2005 at least once a month or two, and don't forget to run the Help Me Stop Smoking Ganjj wizard as needed.
Nonaggression works!