Fork me on GitHub

“You can’t use that, it’s not performant”

I can’t tell you how many times I’ve heard this - from PMs, dev leads, and colleagues.  Other than the use of the “word” performant, what bugs me about this is that usually they don’t have any data (or current data) to back it up.  Coming from the Java world, I’ve probably heard it more than most of my .NET colleagues, since Java is still accused of being slower than X-Windows over a 300baud modem.

However, in the .NET world, this is increasingly used as an excuse to hamstring new development and prevent the use of new features.  Luckily, Vance Morrison has provided folks in the .NET world a way to verify these claims, with his great MeasureIt tool, described in (and available from) this blog post of his.

I recently ran into one of these situations when I was told that I couldn’t use the standard .NET XML Serializer and had to use the XmlReader/Writer classes, because the serializer wasn’t “performant” enough.  This could be a valid concern, but I didn’t believe it – mostly because I didn’t want to write and maintain all of that reader/writer code.  So I grabbed MeasureIt and went to work – I coded up four different tests reading different size XML files with the schema I needed: XmlReader, Xml Serializer, XPath, and LINQ to XML.  These tests were not measuring how well these technologies worked in the general case, but in my specific case, on my specific problem, and with my specific coding irregularities.  I had the tests reviewed by some folks on my team who knew a bit more about .NET + XML than I did, and then went to town.

The graph below measures the time (in normalized time units, on the left) to read different size XML files (on the bottom – the number of high-level objects in the files).

image

Looking at the results, the XmlReader and Serializer implementation look like they perform and scale pretty close to identically.  That’s pretty impressive, and makes me wonder if the Serializer implementation uses XmlReader/Writer under the hood.  I asked around, and it looks like currently that is the way things are done (I’m not on the .NET library team, so I could be wrong), and using sgen.exe with /k I took a look at the generated serialization code and it looks terse.  These results allowed me to justify my use of the serializer – making my code much shorter and easier to test and maintain, at the cost of a very small performance hit.

Performance is important, but it should be viewed as an overall part of the design and development process.  Worrying about performance without having concrete performance goals is a recipe for pain – MeasureIt is a tool to allow you to start these sorts of performance conversations, turning vague performance worries into concrete results that can be discussed practically.  Once you start having these sorts of conversations, you can advance to the point where you have concrete, user-focused and testable performance goals – much easier to satisfy than a vague sense of dread.

Comments !

blogroll

social