Replies to This Discussion

Permalink Reply by Kevin Urban on December 16, 2009 at 2:10pm

My company has created a SQL coprocessor for the analytics market which is used in our DBx appliance. We considered using GPUs for this (which are good with floating point calculations) but we found that GPUs are good at doing only certain tasks, mainly revolving around graphics based calculations. As we were interested in text and integer based SQL operators (group-by, join-by, select project, regular expressions), we decided on using FPGA technology, which is in essence a programmable ASIC. Our results indicate that our coprocessor is about 10x the performance of quad-core CPUs for SQL based operators, and that octal-core CPUs provide only marginal improvements (programming models still aren't taking advantage of core parallelism). We have shown that Monte Carlo simulations are 16x faster than quad-core processors (8x faster the GPUs) running double precision. You can read more about it from our website (which will be updated on Monday).

Permalink Reply by Eduardo on July 11, 2012 at 11:41am

Hello!

Kevin has shows up a good point to us keep in mind. GPU X CPU.

In fact I've never used GPU's for calculation. There are a lot of tech articles pointing for the necessity of re-write programs in order to real get out the GPU power. But several super computers nowadays are betting on GPU's instead of multiples CPU's and getting fantastic benchmarks. I'm looking forward to see what other people have to tell us about!

But Phillip, another important approach is to put memory to work. As much memory you can put in a system, faster they you'll be. Of course, some OS's (Linux/OSX/Unix) handle memory better than others(Win, except servers flavors).

Solid State Disks SSD are far faster than spin disks and could do a great job on data analysis(However, many calibration is required before they came out with the top performance). If we take a look at the new generation of big data servers we'll see lot's of SSD embarked.

Since data bank and data analysis are all about Input/Output - I/O, data read X data write. Minimize this cicle and/or maximize the speed among the several devices as memory, disks, network is a must. Work with faster devices too.

Regards,