In performing the performance tests for Linq vs. ADO.NET, I spent quite a bit of time getting the methodology ironed out. Why? Well, I kept getting different results depending on the order in which the test methods were run. This struck me as somewhat odd and, honestly, even more frustrating. If the methodology was valid, one would certainly expect the results to be consistent regardless of the order in which the test methods were called.

Of course, the first things that comes to mind is the connection pool. The first access to the database with a particular set of credentials would create the pool and take the hit for opening the connection to Sql Server. This would skew the results against the first called test run. This was an easy one and one that I had figured out before even running the tests. Creating and opening the connection before any of the tests were run was a no-brainer.

But something else was going on. The first method called on a particular run seemed to have a performance advantage. I even, at one time on previous tests, had case statements to alter the order ... but even then I'd get different results on different runs. This left me scratching my head a bit. Eventually, though, it occurred to me. There's a bunch of stuff that the Framework does for us and it's sometimes easy to forget about these things and how the impact performance. In this case, it was garbage collection. And it makes complete sense. Think about it ... the GC in non-deterministic. It happens pretty much when the runtime "feels" like it. So ... the GC would happen in various places and invariably skew the results somewhat. The impact didn't seem to be evenly distributed. Why the skewing? Because the GC, when it does a collection, halts all thread processing while it does its thing. Of course, when this occurred to me, it was a "DOH!" moment.

Once I added a call to GC.Collect() after every call to a test method, the results were, as I expected, remarkably similar across all of the test runs, regardless of the order in which they were called. Confirming, of course, my newly realized theory about the garbage collection and its impact on my performance tests.

I did, for the final "numbers" toss out the low and the high values and re-averaged. Since Windows always has other things going on, some of those things may take a time slice or two of the processor from the test run. Or not take any. Still, doing this actually made very little difference to the results. As I think about it, though, I should also create an instance  of every class that I create in order to make sure that the type is initialized in memory and the dll is loaded. But, looking at the results, this really didn't appear to make much difference. Still, on future tests, I'll start doing that.

Now, keep in mind that this applies only to artificial tests. And if you look at the Linq vs. ADO.NET tests, they were certainly quite artificial. Not what you would do in a real-world application. This was, of course, really only designed to test raw numbers for each of the methods that were being used at the time. When you are doing performance testing on your applications, this kind of testing methodology is invalid, to say the least. And calling GC.Collect() after every method call will, without question, hurt the overall performance of your application. So don't do it. For your individual applications, you need to take a holistic approach; test the application in the way it is expected to be used on the real world. Of course, this can only go so far because users will, invariably, do something that we didn't expect (why is that???) and telling them "Well, just don't do that" never seems to be an acceptable answer. For web applications, this needs to go a step further - in web apps, performance != to scalability. They are related, to be sure, but not the same. I've seen web apps that perform pretty well ... but only with a few users, keeling over when they get 20 or more users. That's not good.

I had a little discussion today with an old buddy of mine this morning. I won't mention his name (didn't ask him for permission to) but those of you in Houston probably remember him ... he used to be a Microsoft guy and is probably one of the best developers in town. I have a world of respect for him and his opinion.

So ... it started with this ... he was surprised by the "do you think a user will notice 300 ms".  Of course, that's a loaded question. They won't. But his point was this: 300 ms isn't a lot of time for a user, but under a heavy load, it an be a lot of time for the server. Yes, it can be ... if you have a heavy load. I won't give a blow-by-blow account of the conversation (I can't remember it line for line anyway), but it was certainly interesting.

One thing that we both agreed on that is important for web developers to understand is this: performance is not equal to scalability. They are related. But they are not the same. It is possible (and I've seen it) to create a web app that is really fast for a single user, but dies when you get a few users. Not only have I seen it, but (to be honest here), I've done it ... though, in my defense, it was my first ASP "Classic" application some 10 or 11 years ago; I was enamored with sessions at the time. This was also the days when ADO "Classic" was new and RDO was the more commonly used API. And ... if you are a developer and haven't done something like that ... well, you're either really lucky or you're just not being honest.

With that out of the way ... I'd like to give my viewpoint on this:

Data Readers are still the fastest way to get data for a single pass. If it's one-time-use data that is just thrown away, it's still the way to go. No question. (At least, IMHO). But there's a lot of data out there that isn't a single pass and then toss ... it may be something that you keep around for a while as the user is working on it (which you often see in a Smart Client application) or is shared among multiple users (such as a lookup field that is consistent ... or pretty much consistent ... across all users). In both of these cases, you will need to have an object that can be held in memory and accessed multiple times. If you are doing a Smart Client application, it also needs to be scrollable. Data Readers don't provide this. So ... if you are doing these types of things, the extra 300 ms is actually well worth it, In a web application, you'll scale a lot better (memory is a lot faster than a database query and it keeps load off the database server for little stuff) by caching common lookup lists in the global ASP.NET Cache. One thing that I find interesting ... the LinqDataSource in ASP.NET doesn't have an EnableCaching property like the SqlDataSource. It does, however, have a property StoreOriginalValuesInViewState.  Hmmm ... curious. Storing this in ViewState can have its benefits ... it's a per-page, per-user quasi-cache ... but at the cost of additional data going over the wire (which might be somewhat painful over a 28.8 modem ... yes, some folks still use those). That said, ViewState is compressed to minimize the wire hit and can be signed to prevent tampering. But ... the EnableCaching puts the resulting DataSet (it won't work in DataReader mode) into the global ASP.NET cache ... which, again, is good for things like lookups that really don't change very often, if at all.  For the Smart Client application ... well, DataReaders have limited use there anyway due to the respective natures of DataReaders and Smart Client apps.  Granted, you can use a DataReader and then manually add the results to the control that you want it to display in ... but that can be a lot of code (yeah, ComboBoxes are pretty simple, but a DataGrid ... or a grid of any sort?). One thing that struck me is the coding involved with master/child displays in Smart Client applications. There's two ways that you can do this in ADO.NET: You can get all the parents and children in one shot and load 'em into a DataSet (or object structure) -or- you can retrieve the children "on demand" (as the user requests the child). Each method has it benefits, but I'd typically lean to the on-demand access, especially if we are looking at a lot of data. This involves writing code to deal with the switching of the focus in the parent record and then filling the child. Not something that's all that difficult, but it is still more stuff to write and maintain. With Linq to Sql, this can be configured with the DeferredLoadingAvailable property of the DataConnection and it will do it for you - depending on the value of this property (settable at runtime - you won't see it in the property sheet in the DataContext designer).

There was also some discussion about using Linq vs. rich data objects. This ... hmmm ... well, I'll just give my perspective. This is certainly possible with Linq, though certainly not with anonymous types (see http://blog.microsoft-j.net/2008/04/15/LinqAndAnonymousTypes.aspx for a discussion of them). But ... the Linq to Sql classes are generated as partial classes, so you can add to them to your heart's delight. As well as add methods that hit stored procs that aren't directly tied to a data class.  Additionally, you can certainly use Linq to Sql to have existing (or new) rich data classes that you create independently of your data access and then filled from the results of your query. As for the performance of these ... well, at the current moment, I don't have any numbers but I'd venture to guess that the performance would be comparable to anonymous types.

Performance aside, one thing that you also need to consider when looking to use Linq in your projects is not just the performance, but the other benefits that Linq brings to the table. Things like the ease of sorting and filtering the objects returned by Linq to Sql (or Linq to XML for that matter) using Linq to Objects. There is also the (way cool, IMHO) feature that lets you merge data from two different data sources (i.e. Linq to Sql and Linq to XML) into a single collection of objects or a single object hierarchy. Additional capabilities and functionality of one methodology over another are often overlooked when writing ASP.NET applications ... it's simply easier to look at the raw, single user, single page performance without thinking about the data in the holistic context of the overall application. This is, however, somewhat myopic; you need to keep the overall application context in mind when making technology and architecture decisions. This in mind ... hmmm ... off to do a bit more testing. Not sure if I'll do updates first or Linq sorting and filtering vs. DataViews.

In my last blog post, I took a look at how Linq handles anonymous types. I also promised to do some performance comparisons between Linq and traditional ADO.NET code. Believe it or not, creating a "fair" test is not as easy as one would think, especially when data access is involved. Due to the nature of connection pooling, whichever method is first to be tested gets hit with the cost of creating the connection ... which skews the test. Yeah, I'm sure this is out there in the blogosphere, but I do like to do these things myself. Call it the Not-Invented-Here syndrome.

This particular test set is for a very simple query. I created a set of 4 methods to test for performance within a standard Windows Console Application, which should give an overall comparison of data access. All tests used the AdventureWorks sample database, with the statement (or its Linq equivalent) Select FirstName, LastName From Person.Contact. This is about as simple a query as you can get. From there, each method concatenated the two field results into a single string value ... The Linq test used an anonymous type going against a data class created with the Data Class designer. Data Reader Test 1 (DataReaderIndex) used the strongly-typed DataReader.GetString(index) ... and I did cheat a little with this one by hardcoding the index rather than looking it up before entering the loop (though this is how I'd do it in the "real world"). In previous tests that I've done, I've found that this gives about 10-20% better performance than DataReader[columnName].ToString() ... though that does include the "lookup" that I mentioned previously. Data Reader Test 2 represents the more common pattern that I've seen out there ... using DataReader[columnName].ToString(). Now, I'm not sure which of these methods Data Binding uses and, honestly, that's not in the test ... though, now that I think of it, it may be a good thing to test as well. Finally, I included a test for DataSets (TestDataSet) ... using an untyped DataSet. I've found (again, from previous tests) that this performs far better than a typed DataSet ... the typed DataSet gets hit (hard) by the creation/initialization costs. Before running any tests, I included a method called InitializeConnectionPool, which creates and opens a connection, creates a command with the Sql statement (to cache the access plan), calls ExecuteNonQuery and then exits. This is not included in the results, but is a key part of making sure that the test is as fair as possible. Additionally, all of the tests access the connection string in the same way ... using the application properties. In looking at the code generated by the LinqToSql class, this is how they get the connection string. This ensures that the connection string for all methods is the same, which means that the connection pools will be the same.

To actually do the test, I called each method a total of 30 times from the applications Main, each function in the same loop. This would help to eliminate any variances. After running each test, I also called GC.Collect() to eliminate, as much as possible, the cost of garbage collection from the results.  I also closed all unnecessary processes and refrained from doing anything else to ensure that all possible CPU and memory resources were allocated to the test. One thing that I've noticed from time to time is that it seems to matter the order in which functions are called, so I made a total of 4 runs, each with a different function first. For each run, I tossed out the min and max values and then averaged the rest -- (total - min - max)/(numCalls -2). This gave me a "normalized" value that, I hoped, would provide a fair, apples-to-apples comparison. Each method had a set of 4 values, each with 30 calls, 28 of which were actually included in the normalized value. I then took the average of the 4 values. I know that sounds like an overly complex methodology ... and I agree ... but I've seen some weird things go on and some pretty inconsistent results. That said, in looking at the results, there was not a lot of difference between each of the 4 runs, which makes me feel pretty good about the whole thing.

So ... without further ado ... the results (values are in milliseconds):

Now, I have to say, I was somewhat surprised by the TestDataReaderNoIndex results ... previous tests that I had done didn't show such a big difference between this and TestDataReaderIndex ... though I wonder if that has something to do with the way I did this test - hardcoding the indexes into TestDataReaderIndex. I'm not surprised that TestDataReaderIndex turned out the be the fastest. DataReaders have been, and still are, the absolute fastest way to get data from the database ... that is, if you do it using integer indexes. However, TestLinq didn't come that far behind and was certainly more performant than the untyped DataSet. So ... let's think about this for a second. The Linq collection that is returned is more like a DataSet than it is a DataReader. DataReaders are forward-only, read-only server-side cursors. Use them once and kiss them goodbye. Both the Linq collection and the DataSet allow random access and are re-startable ... and they are both updatable as well. I've had a lot of folks ask about the performance of Linq and now I can, without question and with all confidence, tell them that the performance is quite good.

Still, let's be honest ... the difference between the fastest and the slowest is a mere 300ms. Do you really think users will notice this?

UPDATE: You can download the code and the tests that I used for this at https://code.msdn.microsoft.com/Release/ProjectReleases.aspx?ProjectName=jdotnet&ReleaseId=948. If you get different results, I'd be interested to hear about it. Even more, I'd be interested in the methodology that you used to create the report.

I've been playing with Linq quite a bit recently. I have to say ... it's some cool stuff and revolutionizes data access on the .Net platform. One of the things in Linq that I'm really fascinated with is anonymous types. These classes are created based on a Linq statement and only have the properties that you specified. They're nicely type-safe and work with IntelliSense. Beauty and goodness.

Now, for a time, I just played with them and used them without much thought about what's going on behind the scenes. But ... my curiosity got the better of me and I decided to dig a bit and see what's going on. And the best way to do this? Lutz Roeder's Reflector of course!

So first ... the code. Not much, pretty simple.

            using (DataClasses1DataContext dc = new DataClasses1DataContext())
            {
                var contactNames = from c in dc.Contacts
                                   select new { c.FirstName, c.LastName };
                foreach(var contactName in contactNames)
                {
                    Console.WriteLine(contactName.FirstName + contactName.LastName);
                }
            }

I could have made it even simpler ... remove the foreach loop. But that let's me know that all's well.

So ... what happens with the anonymous type? It's actually compiled in the assembly. Yup, that's right ... it's a compiled class, just like a case that you create. But there is some black voodoo majik going on and, I'm certain, some significant compiler changes to make this happen.

Here's the raw IL generated for the class (with attributes):

.class private auto ansi sealed beforefieldinit lt;<FirstName>j__TPar, <LastName>j__TPar>
    extends [mscorlib]System.Object
{
    .custom instance void [mscorlib]System.Diagnostics.DebuggerDisplayAttribute::.ctor(string) = 
         { string('\\{ FirstName = {FirstName}, LastName = {LastName} }') Type=string('<Anonymous Type>') }
    .custom instance void [mscorlib]System.Runtime.CompilerServices.CompilerGeneratedAttribute::.ctor()

And here's the C# version if the IL:

[DebuggerDisplay(@"\{ FirstName = {FirstName}, LastName = {LastName} }", Type="<Anonymous Type>"), CompilerGenerated]
internal sealed class <>f__AnonymousType0<<FirstName>j__TPar, <LastName>j__TPar>

If you had any doubt at all, the "CompilerGenerated" attribute pretty much says it all. All of the references to the anonymous type in the code are replaced by this class in the compiled IL. And the return value from the query? It's a generic class:
[mscorlib]System.Collections.Generic.IEnumerable`1<class <>f__AnonymousType0`2<string, string>>.

Pretty cool, eh?

Now I'm off to dig into the performance of these beasties when compared to a DataReader and a DataSet. Early results look promising, but I've got some work to do to make sure it's a valid and fair comparison.