Things you may not know.
Machine Learning at POF
Our 147,000 core CUDA GPU Cluster helps us predict who is going to enter into a relationship with who. It also lets us predict and model pretty much anything. We don't use anything off the shelf, and all the code for this is being written in-house.
Technology and scaling is very important at POF. Every month we receive over 10 billion page views, our users exchange 600 million messages, and we serve over 35 billion images. In order to scale to this level we buy a lot of cutting edge hardware. Our Master Database server has 2TB of ram and we use a Ramsan 820
, which allows some of our servers to write over 1 GB/sec off disk. When it comes to database servers, we don't have anything in production that has under 512 GB of ram.
We are well aware that a fast web sites leads to improved user retention which is, of course, a critical factor in our ongoing success. We develop custom solutions and regularly perform low-level inspections of memory dumps to ensure every web page or web service call is delivered as promptly as possible.
We analyze what happens on our website in real time. This is a system that never stops growing and, like most big data projects, we are limited by our ability to log billions of events per day and come up with ways of extracting meaningful data from them. At present, our custom real-time analytics system is processing around 100,000 events every second including web page and web service requests, database transactions, ad impressions, and messaging statistics.
Over half of our traffic comes from mobile devices and most of this from native applications built on iOS and Android. We are constantly making major improvements to these apps, and have new releases every month.
POF users upload images in a wide variety of sizes and formats. These all need to be processed into a common format and several size as well as cached into memory, stored to disk and distributed to a content distribution system. File systems are notoriously slow and with the number of times we need to update and serve our many terabytes of images, we've had to develop new and innovative ways to bypass file system limitations.