Why R Math Functions on Windows are Slow, and How to Fix It

R on windows has much slower versions of the log, sine and cosine functions than are available on other platforms, and this can be a serious performance bottleneck for programs which frequently call these math functions.  The reason for this is that the library R uses to obtain the log function on windows (libmingwex.a) contains a version of the log implementation which is out of date relative to more modern code and much slower than other available versions.  That the glibc implementations of common math functions are slow is a known issue that others have discussed on the internet.  This post uses the log function as an example to show specifically why it is slow, and then suggest some quick work arounds for these math functions.

The log Function on MinGW / Windows

Comparing the assembly code of the log function as generated with MinGW on windows with the assembly generated for the log function on Mac OSX shows why the code is slow. Below is the assembly code for log function as decompiled from MinGW (their implementation is basically a cut/paste from GNU libc). The  crucial feature here is that most instructions start with an “f” indicating that they are using the floating point registers, and there is one instruction the fyl2xp1 that is one of the most expensive operations out there. This instruction takes a log using hardware, and is known to be slower than most other possible ways to calculate log for modern machines. These floating point register instructions were state of the art way back when, but nowadays most hardware is using special XMM/YMM registers and instructions that can make for much faster code to calculate logarithms (See Section 17 for more information on this).

The log Function on Mac OSX

Demonstrating faster assembly code is the Mac OSX implementation of log shown below.  The key feature here is that it uses XMM registers and has a more modern and performant implementation.  Note that many other implementations (e.g. the Microsoft compiler) also use faster versions like this. That is, Windows is not slow, R is slow on Windows. It’s all the same Intel chips under the hood.

Solving the problem

Ideally we could just push a faster log implementation to the R core repo for Windows, but the fact of the matter is the R core team does not accept code changes that only improve performance (and fair play to them, R is the most cross platform software out their, and that wasn’t easy to do).  Also, realistically the current implementation is feasible for most peoples work. So the solution is to switch out the log function for a better one only when it matters, and if it matters that means most of the library is already written in compiled code. Where to get a cross-platform compatible log function? It seems the folks over at Julia ran into a similar problem with slow Log on different machines. They looked at two version in libm (e.g.  Libm Version #1 and Libm Version #2), and also some really crazy implementations such as this one available from Yepp. However, what they settled on is a very good function, and we can port those over directly into R.  An example is shown in the pull request here.    

Profiling Rcpp package code on Windows

Profiling Rcpp code on Unix/Mac is easy, but is difficult on Windows because R uses a compilation toolchain (MinGW) that produces files that are not understood by common Windows profiling programs.  Additionally, the R build process often removes symbols which allow profilers to produce sensible interpretations of their data. The following steps allow one to profile Rcpp code on windows.

Change compilation settings to add in symbol settings

A default R installation typically has certain compiler settings placed in the equivalent of the C:\Program Files\R\R-3.3.1\etc\x64\Makeconf that strips information needed for profiling during the Rcpp compilation process, in particular a line which reads:  DLLFLAGS=-s . To override this and add some additionally needed flags, one should add a folder and file to their home directory which overrides and appends necessesary compilation flags.  To a file located at a location equivalent to  C:\Users\YOURNAME\.R\Makevars on your machine (note the ‘.’ before R), add the following lines: You can verify this worked correctly by checking that -gdwarf-2 appears in the compilation messages, and that -s is missing in the final linker step.

Run a profiler which understands MinGW compiled code

The next key step is to run a profiler which can understand the Unix like symbols on windows.  Two free and good options are Very Sleepy and AMD’s code analyst (which also works on Intel chips).  Very Sleepy is very good at basic timings and providing stack traces, while AMD’s profiler is able to drill down to the assembly of a process. Both profilers are good but an example with AMD is shown below.
  1. Open the program and setup a quick session to start and run a sample R script that uses your code, such as in the example shown below.AMD_ProfilerSettings
  2. Next run the profiler and get ready to look at results.  For example, here I can see that half the time was spent in my code, versus half in the R core’s code (generating random numbers)ProfilerResults1And digging further down I can see at the assembly level what the biggest bottlenecks were in my code
assembly Its often helpful to look at the original source files in addition to the assembly, and this can be enabled by setting directory information by Tools-> CodeAnalyst Options -> Directories.  

C# vs. Java, Xamarin vs. Oracle, Performance Comparison version 2.0

Today I noticed the SIMD implementation of the Mandelbrot set algorithm I blogged about last year was successfully submitted to the language shootout webpage. However, I was a bit disappointed to see the C# version was still slower than the Java version, despite my use of the special SIMD instructions (though I was pleased that a quick port of my code to F# absolutely annihilates the OCAML version of this benchmark).

I had benchmarked my code as faster than the Java code on my Mac, the CentOS cluster at my work and two Ubuntu virtual machines (Mono >3.0, LLVM compiler). What gave?

Undeterred, I thought I would try to use SIMD to improve another C# benchmark, and took a shot at the N-body benchmark test. Once again, the version I wrote, in my hands, was much faster. But when submitted to the shoot-out and accepted, it was either much slower, or didn’t even run. My submission using the SSE instructions not only didn’t beat Java, it was actually slower than the original C# version!

Below are the timings on my top-of-the-line Mac for the two benchmarks, in both cases we see that the C# program runs in 80-90% of the time the Java Program takes. There are several key take aways here.


C# and Java have no meaningful performance differences.

C# may use a lot less memory, but this is my optimized C# code and it is only beating the optimized Java by <= 20%. When does that matter?

C# and Java’s similar performance has different underpinnings

There are a lot of ways code can be fast, and I was surprised that Java and C# achieve similar speeds in very different ways.

The key advantage of the C# code was the SIMD instructions, which in theory give a ~2X speed up. However, they only win by ~20%. Why?

I think the assembly gives the answer. Just some quick snippets tell the whole story. Here is some C# code for the N-Body problem compiled to assembly:

The important insights into performance from the assembly here are:

  1. Similar instructions are stacked on top of each other, allowing for pipelining (e.g. the same vhaddpd instruction follows the same vhaddpd, and both use different registers, so can execute simultaneously).
  2. The “p” in the instructions (i.e. vhadd-“p”-d). This stands for “packed” meaning we are packing/doing two operations at once via SIMD.
  3. Only registers XMM1-XMM4 appear in the instructions. There are more registers available, and more pipelining possible, but the Mono/LLVM compiler appears to only use the low-number/scratch registers. It is easier to write compilers that obey this restriction.

Now let’s compare that to some Java assembly emitted by the Oracle runtime for the same benchmark:

The important insights here are:

  1. Java uses the version of the instructions with “s” (i.e. vmul-“s”-d) meaning single, and gets no SSE advantage. No doubt this makes writing compilers easier.
  2. However, Java is using lots of registers, XMM15 shows up!
  3. As a result of using all the registers used, the Java code has great pipelining. Note that up to 5 vmulsd instructions show up at once. The JVM is simply brilliant at packing operations together, and this means that even though I, the programmer, was smart and used SSE2, my C# code only won by 20%, and not 200%. It’s hard to beat Java pipelining.

All of which makes me wonder. What if we took the high-level language advantages of C# (low-overhead value types, easy interop via pointers, baked in generics, etc.). And combined them with the low-level advantages of the JVM (better array bounds check elimination, pipelining compiler optimizations, and maybe someday the occasional stack allocation of reference type variables…)

NuMTs, mtDNA sequencing and Aligners

There are a lot of NuMTs (nuclear encoded mitochondrial sequences) in the genome, and when the mtDNA is sequenced, so reads may align to the nuclear genome instead of the mtDNA because of this.  But how much winds up in the nuclear DNA and where does it go? To answer this, I simulated reads from a diverse collection of mitochodria, and tracked where they landed when aligned with bwa mem. The reads were simulated from the whole collection of mtDNA molecules available from phylotree, and the simulated reads were 100 bp in length, have a 1% error rate, and an insert size normally distributed around a mean of 150 bp with a std. dev. of 30 (but bounded at a minimum of 40 insert and max of 700). After simulating, I then aligned with.   And discovered that almost all reads align to the mtDNA, only 3% of reads aligned elsewhere.  As a result, the distribution of coverage depth across the whole genome is very bi-modal. Histograms showing the coverage depth distrbution of sites with data is shown below. img5 For reads that did align to the nucleus, the MAPQ was typically 0, but could be as high as 60 and had an unexplained peak at 27. img15 And below shows the normalized coverage by positions across the mtDNA, clearly some regions are more affected by NuMTs. imgD Reads from the first and last 500 bp of the mtDNA are poorly aligned by bwa.  It appears most go to chromosome 17, but their true location is belied by their mate pair. In fact only 0.6% of reads in this region that map to the nuclear DNA do not have their paired read map to the mtDNA. img19 I also wanted to see how reads that represent a heteroplasmic deletion would be handled.  I simulated reads that either spanned or included a deletion randomly chosen to be in the mtDNA, again virtually all mapped to the mitochondria, and the coverage profile looked similar to the simulation with complete reads. Perhaps most reassuringly, almost all reads are mapped. Checking for unmapped reads with the command: Showed only one un-aligned read out of the simulated millions, and this read had many errors compared with the original sequence it was simulated from. The result of all of this is one large BedFile giving the location of all possible reads from elsewhere.

The .NET Bio BAM Parser is Smoking Fast

The .NET Bio library has an improved version of it’s BAM file parser, which makes it significantly faster and easily competitive with the current standard C coded SAMTools for obtaining sequencing data and working with it. The chart below compares the time it takes in seconds for the old version of the parser and the current version to parse a 76 MB BAM file. The current parser can easily create ~150K validated sequence objects per second on the clunky old servers I typically run code on. Note that the windows and unix numbers are from completely different machines and not comparable. Also included is a comparison to a “custom” version of the parser that I wrote, which uses unsafe code, assumes the system architecture is always little endian, caches strings and does some other tricks to get some further performance improvements and reduce memory usage. img5 The comparison to samtools is based on the system time to parse the file using this task on the same unix server used for the C# tests. samtools view Test.bam | wc -l And is just meant to give a general idea of the performance comparison, as there are several important differences between the samtools and .NET Bio test. The C# version was not allowed to reuse memory for objects as it was supposed to be working as a data producer, while the Samtools version processes reads one at a time and does reuse memory. C# also made a lot of dictionaries to aid quick access to the read groups, which isn’t done by samtools. However, samtools had to write the files to the output pipe, while the C# version did not, which undoubtably introduces a fair bit of overhead for it. Both tools however, are clearly plenty fast and at this stage further performance improvements would come from lazy evaluation (or not sticking unnecessary information like the original quality scores in the BAM files!), and the language won’t matter much. Performance Comments One task when parsing BAMs is unpacking lots of information that is packed together in arrays.  In SAMtools and the current .NET Bio parser, this is done with lots of unpacking of bits by integer manipulations.  For example this code from SAMTools: Because C# has pointers and value-type structs however, I discovered that it is a lot more fun just to define a structure that contains those fields and unpack directly with a pointer cast in C#. Blam! Now all the data related to the position, bin read group is in the object with those three lines that copy the data very fast. So where are the bottlenecks remaining? On windows about a third of the time is spent doing the decompression. In Mono, because the decompression is done by zlib and not in managed code, it’s effectively free. Currently, the quality data and sequence data are passed around a bunch, and the code could likely be made about 10% faster by not copying that data but reusing a single byte array each time. However, it is so fast it hardly seems worth worrying about.

Using Selectome with .NET Bio, F# and R

The Bio.Selectome namespace has features to query Selectome.Selectome is a database that merges data from Ensembl and the programs in PAML used to compute the ratio of non-synonymous to synonymous (dN/dS) mutations along various branches of the phylogenetic tree. A low dN/dS ratio indicates that the protein sequence is under strong selective constraint, while a high one indicates that selective constraint is more relaxed. Selectome is also a fantastic resource to get gene trees and multiple sequence alignments. Using selectome and .NET Bio allows you to quickly investigate divergence across the vertebrate phylogeny. This page gives a walk of how to query selectome, convert the text data it returns in to objects, and the compute and plot various quantities from those objects.

Example Walk Through.

Step 0: Setup F#–If you haven’t used F# before, you can download Tsunami. Open the program and add references to Bio.Selectome using the #r command and open the relevant namespaces. (highlight and hit Alt+Enter to run in the console).

Step 1: Make Gene List – Selectome requires ensembl identifiers be used in queries.To create a set of interesting genes, I first downloaded the full set of genes from the MitoCarta website. These genes are identified by Entrez IDs, while selectome uses Ensembl IDs, so to convert between these I used the GeneID converter website to create a flatfile of the new ids. Given this flatfile, we can load it and convert it to typed classes as follows:
Step 2: Query Selectome –  All selectome data is accessed through the SelectomeDataFetcher class. This class will return a SelelctomeQueryResult that will let you know if the query was successful. Currently, the queries will only be successful for genes that exist in the database and have data available for the full vertebrate tree. If no data is available the Result will be NoResultsFound, if selectome returned data but there was no tree available for all vertebrates(but maybe just primates) the result will be NoVeterbrateTreeDataFound. We want to extract genes from query results that successfully returned data for the vertebrate tree.
Step 3: Determine how many genes show positive selection – F# makes this easy: Interestingly, roughly 33% of genes show selection, so we know not to get too excited by any one result!

Step 4: Download Multiple Sequence Alignments  – In order to decide how conserved a protein is relative to other proteins, we can download the multiple sequence alignment for each protein in this set and compare it to a particular protein of interest.  In Selectome, each protein comes with a masked and unmasked alignment for both proteins and DNA. These objects are available from the SelectomeGene class and are lazily downloaded when requested from the Selectome server.  These sequence alignment downloads are also cached for 30 days in a temporary directory to avoid waiting for downloads if you want to reload your script of interest.  Once downloaded they are converted to full-fledged .NET BIO multiple sequence alignments, meaning one can do nearly anything with them. The example below gets the DNA alignment and the BLOSUM90 alignment score for the masked amino acid alignments.

Step 5: Download the Gene Trees – The selectome gene defines a class, SelectomeTree, that provides a great set of functions to query all the interesting metadata provided by selectome. These features are most usefully investigated by using the autocomplete functionality of your editor, but there is a lot of useful information! Some examples are also shown below.

Tree queries are also cached locally to avoid going back to the server in the event of repeated requests.
Step 6: Plot distribution of interest using the R data provider – You can call R plotting functions directly from F# using the R type provider. More information is available from that site, but the code snippet below is sufficient to produce a histogram of alignment scores, no need to dump to a flat file first! Huzzah! One intermediate machine to rule them all (or at least to avoid useless glue between different libraries/APIs).

Mono.Simd and the Mandlebrot Set.

C# and .NET are some of the fastest high level languages, but still cannot truly compete with C/C++ for low level speed, and C# code can be anywhere from 20%-300% slower. This is despite the fact that the C# compiler often gets as much information about a method as the C compiler.  It has been suggested that SSE/SIMD instructions in C# could overcome some of these issues.  To evaluate this, I took a famous computational task, re-coded it using C# SIMD instructions, and evaluated the performance by looking at the execution time and how the emitted assembly code compared to the optimal assembly code.


In 2009, Miguel de Icaza demonstrated a framework that allows C# to use SSE2 intrinsic operations.This is now part of the Mono library and in theory, such methods can greatly save computational time (1.5X-16X), particularly for operations on byte arrays, a common task in bioinformatics.

Test Case

Computing the mandelbrot set is one of the tasks of the computer benchmarks game, which compares program speed in different languages.  Currently, C# is listed as being slower than Java, though the programs in each language use different techniques and neither uses an optimal algorithm (see here for a better one).  However, it makes a useful test case for raw computing speed.  The challenge is to compute the picture shown below by recursively squaring a complex number and adding a constant to it.

z_{n+1} = z_n^2 + c


The Algorithm

Complex numbers are a particularly challenging case because their multiplication is not a simple operation, but involves a rather annoyingly squaring of different terms and then adding/subtracting them.


These are not easy vector operations, and as will be shown later one result of this is that using SSE to speed up the values for one multiplication is useless (I tried, it was worse). So, the performance improvement comes from doing two elements of the loop at a time (this is how the current Java submission is faster than the C# one, though it does not use SIMD).  The current C# version does the inner loop of the program without SIMD instructions as follows:

This loop is iterating until convergence (<4) or until the max iterations (i<0) have expired. See the wikipedia page for a full explanation.  Of interest here is that the “t” variables exist for pipelining purposes, so this is a reasonably optimized code snippet.

Trying to Implement the Optimal Solution Using Mono.Simd

Actually figuring out how to do the complex arthimetic with SSE2 is rather challenging.  Fortunately Intel published a document giving the best solution as hand coded assembly, which involves using some special SSE3 instructions I was not aware of.  Notably, the Intel optimized code is far faster than even their C code, but is only about 25% better than their optimized assembly code without SSE3.

The code below shows my attempt to implement the Intel approach in Mono.  It should be pretty fast, however it may suffer from one fatal flaw.  The comparison at the end to check for the final conditions currently requires unpacking both values in the Vector2D (when either Length.X or Length.Y have been < 4.0 at least once, then the loop stops).  The comparison for both X and Y can be done in one operation in SIMD using the built in less than statement.  However, I do not know how to turn that into a control loop function, as this requires a movmskps assembly function which mono does not seem to expose.

Results – It’s Faster on Ubuntu


Shown above are the times for three different algorithms. The first is the original best submission on the website.  My SIMD submission is shown on the bottom.  Because the SIMD version does two cycles within each inner loop, as the Java submission does, I tested the Java submission converted to C# as well.  Compared to the current submission this version shows a 81% improvement, but clearly much of that is simply from doing 2 cycles in one loop.  However, even once this change is made, the SIMD instructions still give a performance boost.

The Assembly Generated is Still Not Optimal

Although the assembly generated did include lots of SSE commands, in general inspecting the assembly I noticed several things.

  1. Never unpack the double array values!  My first attempt tried to do the SSE2 steps with those instructions, and then unpack the single values as needed.  However, this failed prettty miserably, as accessing the double seemed to involve a lot of stack to register movement.

  2. All the XMM registers are not used.  The optimal version uses all of them, the C# version uses only 0-3, and soften moves things to and from the stack. Not using registers optimally seems to be a common problem though with C#.


The SIMD version was indeed a fair bit faster, which is nice!  However, in this case it was not a game-changer.  Most importantly though, it was really easy to use, and I think I might incorporate the byte array operations at some point in future code.  This also gave me an appreciation for assembly, which in contrast to what I had heard is easy to read and seems easy to optimize.  I just submitted the code to the shoot-out, assuming it works there it should be up soon, and I would love for someone to show me how to fix the end statement.

Accessing dbSNP with C# and the .NET Platform

NCBI Entrez can be accessed with many different platforms (python, R, etc.) , but I find .NET one of the best because the static typing makes it easy to infer what all the datafields mean, and navigate the data with much greater ease.

Documentation is sparse for this task, but here is how to access NCBI from the .NET platform.  The example steps/program show how to query dbSNP for information about particular ids.

  1. Open visual studio, start a new project and add two Service References to the project: http://eutils.ncbi.nlm.nih.gov/soap/v2.0/eutils.wsdl and http://eutils.ncbi.nlm.nih.gov/soap/v2.0/efetch_snp.wsdl files. Note that the efetch depends on the database used, in this case “snp” other databases have different ones.  You should now have two references:image

More information is available here on setting up visual studio: http://www.ncbi.nlm.nih.gov/books/NBK55694/

2. Next up grab the data as shown in the example.  Each SNP has A LOT more data, which can be inspected in the IDE.

   1:  static void Main(string[] args)
   2:  {        
   3:          string dbSNPid="28357684";
   4:          efetch.eFetchRequest re = new efetch.eFetchRequest();
   5:          re.id = dbSNPid;
   6:          var serv=new efetch.eUtilsServiceSoapClient();
   7:          var exchange = serv.run_eFetch(re);
   8:          var snpData = exchange.ExchangeSet.Rs[0];
   9:          object[] dataToReport=new object[] {
  10:              snpData.Het.value,
  11:              snpData.hgvs.First(),
  12:              snpData.PrimarySequence.First().accession,
  13:          };
  14:          Console.WriteLine( String.Join("\t",dataToReport.Select(x=>x.ToString()).ToArray())); 
  15:          Console.ReadLine();
  16:  }

The following links contain other useful information for consuming the entrez webservice in C#.

Setting up visual studio: http://www.ncbi.nlm.nih.gov/books/NBK55694/

Using efetch: http://www.ncbi.nlm.nih.gov/books/NBK55694/#csharp_msvs.Call_HYPERLINK__chaptercha_8

Forming queries: http://www.biostars.org/p/3436/

More information: http://eutils.ncbi.nlm.nih.gov/entrez/eutils/soap/v2.0/DOC/esoap_help.html

Also note that the first query takes much longer than subsequent ones, for reasons unknown to me at present.

Java vs. C# Performance Comparison for Parsing VCF Files

Making a comparison with a reasonably complex program ported between the two languages.

Update 3/10/2014: After writing this post I changed the C# parser to remove an extra List<> allocation in the C# code that was not in the Java code.  After this, the Java/C# versions are indistinguishable on speed, but the C# code used ~88 MB of memory while the java version used >1GB.  Therefore, I now believe the winner is C# and a fast implementation of this parser (which can be over an order of magnitude faster for certain scenarios not in this test) is available here

VCF files are a popular way to store information about genotypic variation from next generation sequencing studies.  The files are essentially large matrices, where each element in the matrix represents a collection of information about the genotype of a particular person at a particular locus in the genome (in this sense, they can be considered as a multi-dimensional matrix in a flat file format). The Java Picard package is a common utility used for parsing these files.  While parsing, the challenge is to read each line (row) of the file, and construct objects for each element in that row that can then be manipulated or inspected.  I just finished translating the Java VCF parser in Picard to C#, and so thought it might be a good chance to compare the two different languages and runtimes. C# showed a number of advantages in the translation.  The translation itself was mostly a lot of deleting.  The get/set assessors in C# really allowed for the removal of seemingly endless amount of getXXXX/setXXX methods in Java.  It also seemed like every other line in Java was a call to some apache commons class to perform a simple task like get the maximum value in an array, create an empty list, or do a selection on data.  Extension methods and Linq have clear advantages for data processing here (though I have found these have a slight overhead relative to the for loop equivalents).  Yield statements in Java also would seem to be useful. At the same time, Java had some things that would have been nice in C#.  I had to implement basic collection types like immutable hashsets and dictionaries, a LinkedHashSet class as well as an OrderedGenericDictionary during the port.  These should be in the C# language.


This of course am what I am most interested in.  My main computer is broken, so I had to test on my windows desktop at home.  For the test, each program would read a gzipped VCF file for 20,000 lines, first creating an initial lazy class representing a portion of the data and then fully decoding this lazy version to represent the complete data as objects.  The test file was a VCF with >1,000 individuals, though unfortunately most of these individuals were non-calls at various positions, but its what I had on hand. Immediately after porting – After essentially re-writing the Java in C#, I ran some tests.  Both Java and C# can run in either client (low-memory) or server modes.  So I did both, here are the results:
Platform VM Options Working Set Paged Memory Time (s)
Java None 27.8 MB 41.32 MB 11.5
.NET None 28.9 MB 30.22 MB 15.1
Ratio 1 1.35 0.76
Java Server 362.2 MB 414 MB 7.4
.NET Server (GC) 126 MB 332 MB 14.7
Ratio 2.9 1.25 0.5
A couple noticeable conclusions here. First Java is smoking .NET on performance, but this is essentially Java code at this point and it wasn’t written for C#.  Second, there is a massive amount more memory used in server mode, and in Java at least one obtains a large performance win for this cost. After “Sharpening” the code – The initial port was basically Java, so I cleaned it up a bit after running it through the profiler, here are some notable changes:
  • String.Split winds up being a big cost in this parsing.  When porting I initially just recreated an array every time, after I realized I was recreating such large arrays I reused them as in the Java code.
  • In C# you can use unsafe pointers and I got a big performance win out of these.  I grabbed the System.String.Split code from the Mono framework, trimmed/altered it, and used it for all the split methods that seemed to be taking a long time.  The Java version also implements a custom Split method, though obviously can’t use pointers.
  • Some additional cleanup in the logic.
Platform VM Options Working Set Paged Memory Time (s)
Java None 27.8 MB 41.32 MB 11.6
.NET None 28.9 MB 30.22 MB 12.6
Ratio 1 1.35 0.92
Java Server 362.2 MB 414 MB 7.4
.NET Server (GC) 123 MB 181.61 MB 10.8
Ratio 3 2.27 0.68
So round 2, and once again Java is the winner for speed in server mode, though at a high cost in memory.  For the lower memory client model, it is nearly a tie between the two.

What explains the difference?

These programs are nearly identical but the bottleneck is not the same in both. It seems C# can split strings much faster and Java can allocate memory much faster.  Although I am less familiar with the Java profiler, it seems to show 66% of the time is spent on the String splits.  In contrast, in C# the methods that are taking up time have to do with allocating memory (constructors) and the GC, string splits are only ~17% of the total time. image image One the one hand, this means that the JVM is really doing a great job in server mode on memory allocations and other optimizations.  I can’t think why C# shouldn’t be able to match the JVM performance (perhaps dynamic recompilation is really killing it here).  On the other hand, it means that the C# program can still be improved, while I can’t really see how to improve the Java one.  The String.Split method has already been rewritten, and I didn’t see any reasonable improvements for it.  In contrast, both programs have several places where memory allocations can be saved.  For example, one aspect of the parser is that it relies on a factory class to create another class and so allocates twice the memory that it needs as it creates two large objects.  Simply having one object be it’s own factory would solve this.  Similarly, several empty lists are repeatedly created that could point to the same list or simply be skipped.  I wanted this parser to generally match the Java version, so did not pursue these changes, but my guess is they may shrink the difference (though again, in Java this clearly didn’t matter).


The JVM is the clear winner on speed, particularly in server mode where memory isn’t an issue.  C# was the clear winner on brevity, syntax and language features.  The difference was only substantial in server mode, and the C# program (and likely the Java program) were far from optimal, but it gives a rough hint at how they compare.  The next question will be how they compare when C# runs with mono on a Linux environment.

Software Patents

In the dark moments when one wonders if their research is doing enough to better the world, it’s always nice to remember you’re not a patent lawyer.  The radio program This American Life did a fantastic investigation of this issue recently, highly recommended and very entertaining, so I thought I would point it out here.  It explains how broken the system is, and although I don’t know how best to ensure people are compensated for their creative works, clearly we need better solutions.

The outcome of perceived or real patent fights basically moved me away from the Linux desktop.  In 2006 I thought the Linux desktop would become the world standard (the other options were far from awesome).  Instead the technology company that spends the least on research and development rose to prominence and essentially everyone I know who uses Linux comes from the tail end of the most affluent class of society (which probably makes them great targets for the now integrated Amazon search).  Here’s hoping this mess gets cleaned up in the next few years.