Rebuilding the Enterprise - Software, Hardware and Peopleware Migrations for the Systems Architect


Performance Tuning During Development

6/05/2014 Posted by William Berry , , , No comments
The day (and lately, night) job has been focused on a data api backing a new business insights tool.  Despite concerted efforts by nefarious interests, we have managed to keep the API resource centric, only leaking behavior where absolutely necessary.  We are starting the process of putting internal customers in front of the tool, and as such, I have become increasingly aware of the performance issues, of which there are quite a few.

The API provides an abstraction over two different persistence store technologies (Sql Server & HBase) and the domain models are built from data that resides in both stores.  No one particular resource channel is slow, but in the aggregate it *could* be more performant.  We have been working hard over the last few days to weave instrumentation into the application using Semantic Logging, and I figured, while I was down in bowels, I could investigate what I perceived to be performance problems.  

I started with the usual Visual Studio performance tooling ... fire up Performance Analyzer and run all of the stock wizard tests one by one (except for the contention test - I am not convinced it ever really gives useful information).  These tests should be focused and limited in scope.  Running the analyzer for 20 min and exploring every corner of your API is for another time.  Here we are just trying to keep tabs on sticky or overly complex code blocks and you really just need to scout for a few key data points:
  • What methods in your stack or in the core libs are called most often? What is their percentage of total calls?
  • What blocks of code are flagged in the hotlines graph?
  • Where is most of the processing time spent?
  • What spends the most time waiting for other methods to finish?
Don't spend more than a few min just skimming the reports, remembering that you are here purely to simplify data structures, make your code cleaner and maybe a bit faster.  

Below is a classic example of scaffolded code that I started a new algorithm off with.  It was filled in with business logic, and then became a blight on the system.

First the scaffold code:

  1 var dedupedEdges = new List<Tuple<string, string>>(); 
  2 foreach (var edge in edgeGroups.SelectMany(x => x.EdgeList))
  3     // Make sure both points are not "green" 
  4     // Tuple is not in list as (item1, item2) 
  5     // Tuple is not in list as (item2, item1) 
  6     { 
  7     dedupedEdges.Add(new Tuple<string, string>("","")); 
  8     }

After writing a few failing tests we have a naive implementation ... that works perfect over the small data set we are unit testing against.

The code as first implemented:

  1 var dedupedEdges = new List<Tuple<string, string>>(); 
  2 foreach (var edge in edgeGroups.SelectMany(x => x.EdgeList)
  3     .Where(x => !x.Item1.IsGreen() || !x.Item2.IsGreen()) 
  4     .Where(edge => !dedupedEdges 
  5         .Contains(new Tuple<string, string>(edge.Item1, edge.Item2))) 
  6     .Where(edge => !dedupedEdges 
  7         .Contains(new Tuple<string, string>(edge.Item2, edge.Item1)))) 
  8     { 
  9     dedupedEdges.Add(new Tuple<string, string>(edge.Item1, edge.Item2)); 
 10     }

Now I am not proud of this code; it's repetitive and poorly designed.  But, it's also done, working and passing all the tests.  Don't fool yourself, if you write enough code, day in and day out ... a non-neglible percentage is pure drivel like this.

So I am merrily flipping through the reports only to find that I am making 1.25 million ... yes I said MILLION calls to Tuple.Item1 and another 1.25 million calls to Tuple.Item2 during one very large API call.  So following the hotlines graph, I am kindly directed by Visual Studio to the crap starting on line 2.  Knowing better than to just start hacking away at it, as my instincts directed, I put in some code to help me understand the severity of the problem first.  "If it ain't Baroque don't fix it"

Stopwatch() ... as the guys on DotNetRocks say "know it, learn it, love it."  I dropped one around this  block with a handy little Debug.Print(Stopwatch.EllapsedMiliseconds) and found my 2.5 million calls to Tuple.Item* were taking 1753ms, over an average of 10 runs across this code.  First, the .Net framework can clearly make very, very bad code run not so badly.  Second, I clearly have a demonstrable problem here.

When I wrote phoned in this code, I was focused on the bigger algorithm at hand ... "pass the tests and move on."  But now, armed with metrics, I have justifiable cause to go a'hacking.

There are a number of things that need to be fixed, but let's first take a look at what this code is supposed to do.
  • The usage of SelectMany indicates in this context that we are iterating a child list of multiple EdgeGroups.
  • An Edge is a Tuple<string & string>.
  • The first predicate filter makes sure both elements are not "green" (whatever green means).
  • We then check the dedupedEdges List to make sure a new version of the Tuple is not already in the list.
  • We next check the dedupedEdges List to make sure the reversed Tuple is not in the list.
  • Finally take the Tuple, decompose it, rebuild a new tuple and finally add it to the dedupedEdges List.
What, pray tell, could possibly be wrong here?  Let's go through each step and see if we can't make this better, shall we?

Well the first step is to not use a crappy data structure.  What we need here is a Set, someplace where we can only stick 1 of an item and not end up with any duplicates.  Additionally, we need to do fast lookups into the data structure to make sure the reversed tuple is not present.  Hello, HashSet.  Let's see what this code looks like after a refactor:

  1 var dedupedEdges = new HashSet<Tuple<string, string>>(); 
  2 foreach (var edge in edgeGroups.SelectMany(x => x.EdgeList)
  3     .Where(x => !x.Item1.IsGreen() || !x.Item2.IsGreen()) 
  4     .Where(edge => !idIncludedEdges.Contains( 
  5         new Tuple<string, string>(edge.Item2, edge.Item1)))) 
  6     { 
  7     dedupedEdges.Add(edge); 
  8     }

The loop still does the original predicate on green-ness; but now, we are doing do a highly performant lookup on the reverse tuple construct and finish by blindly tossing the filtered items at the HashSet which will ensure uniqueness for us.

Is this perfect?  Nope.
Could it be better?  Sure.
Do I care? Nope.

I get paid to write code ... hopefully, a lot of code.  If I can produce that code quickly great!  If that code is quick itself, greater sill!  If someone else can understand my quickly generated, quick code - I can stop for the day, as I have succeeded.

So, the totally unscientific performance boost from this refactor ... 1753ms ... to an average of 5ms over 10 iterations.

Core Takeaway:

  • You can do "some" performance work during development without hindering forward progress.
  • If you chose to do performance during active development, keep your efforts reasonable.
  • Deep dives while you are still developing an algorithm are distracting.
  • Often you can make huge performance gains without changing the shape of the code.
  • You will write crappy code and 90% of the time ... thats OK.
  • Know that you write crappy code, own it, hunt for it, fix it.


Post a Comment