Subject: [R] Kudos for R team Date: Sat, 21 Sep 2002 21:22:29 -0400 From: "Vele Samak" To: Hi, I just wanted to give kudos to the R team for building a performance oriented software for statistics. As someone who deals regularly with tons of data speed in analysis is a priority. I'll mention one example where the implementation of R and Splus differs so much that leaves you no choice but to go with R. I my line of work, quantitative investment research, we often need to create lots of different but realistic portfolios for backtesting of our models. One function that we developed takes an existing portfolio and generates N various portfolios which are similar to the starting portfolio and meet certain constraints. The function is all S code and uses couple of sapplys, a while loop at the core and some calls to runif. Nothing special. It returns a matrix S x N (S is number of stocks in the starting portfolio). We ran this function on the latest Splus, W2000, pIII 1GZ, 512MB RAM, with e 500-stock portfolio (vector) and asked it to give us 2000 portfolios. It took close to 1 hour. Same code, same machine no changes, R 1.3.1 and R 1.5.1: under 1 minute!!! This isn't geeky fast, fast for bragging rights fast, 10%, 20% or even 100% faster. This is an order of magnitude faster. All in the sapply and the while loop implementations, probably some cbinds in there, but you can speculate. At this point, you have to wonder what goes on in Splus. This kind of result also settles the question once and for all: nor why R, but why splus? No reason at all. It used to be that I had a variety of reasons to use R over splus for the past 2 years, but this example outweights all other. Lest I get fired, I can't show you the code of the function since it is considered proprietary. Although, I doubt you need much time and tought to replicate something like this in S. At this point we moved the core into c and run such tests with all kinds of group constraints in mere seconds. Maybe most of you knew about this, but I just wanted to put this out there. Another side issue, in case there are other believers in Sun workstations out there: R 1.3.1 compiled on Sun 4-cpu workstation with 4GB RAM, don't know the model, but sure it's no more than 2 years old, late 2000 model: same code, same process, mostly lots of regressions, minor I/O to disk: 7-times slower than same execution on the abovementioned PC. Go figure! P.S. Don't want to start flame wars, I but don't make these statements lightly, this is what I get paid to do and every resource and every second in execution counts. Enjoy, -- Vele Samak http://www.velesamak.com -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-help-request@stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._