System Administration of the IBM Watson Supercomputer
System administrators at the USENIX LISA 2011 conference (LISA is a great system administration conference, by the way) in Boston in December got to hear Michael Perrone's presentation "What Is Watson?"
Michael Perrone is the Manager of Multicore Computing from the IBM T.J. Watson Research Center. The entire presentation (slides, video and MP3) is available on the USENIX Web site, and if you really want to understand how Watson works under the hood, take an hour to listen to Michael's talk (and the sysadmin Q&A at the end).
I approached Michael after his talk and asked if there was a sysadmin on his team who would be willing to answer some questions about handling Watson's system administration, and after a brief introduction to Watson, I include our conversation below.
What Is Watson?
In a nutshell, Watson is an impressive demonstration of the current state of the art in artificial intelligence: a computer's ability to answer questions posed in natural language (text or speech) correctly.
Watson came out of the IBM DeepQA Project and is an application of DeepQA tuned specifically to Jeopardy (a US TV trivia game show). The "QA" in DeepQA stands for Question Answering, which means the computer can answer your questions, spoken in a human language (starting with English). The "Deep" in DeepQA means the computer is able to analyze deeply enough to handle natural language text and speech successfully. Because natural language is unstructured, deep analysis is required to interpret it correctly.
It demonstrates (in a popular format) a computer's capability to interface with us using natural language, to "understand" and answer questions correctly by quickly searching a vast sea of data and correctly picking out the vital facts that answer the question.
Watson is thousands of algorithms running on thousands of cores using terabytes of memory, driving teraflops of CPU operations to deliver an answer to a natural language question in less than five seconds. It is an exciting feat of technology, and it's just a taste of what's to come.
IBM's goal for the DeepQA Project is to drive automatic Question Answering technology to a point where it clearly and consistently rivals the best human performance.
Watson's Vital Statistics
90 IBM Power 750 servers (plus additional I/O, network and cluster controller nodes).
80 trillion operations per second (teraflops).
Watson's corpus size was 400 terabytes of data—encyclopedias, databases and so on. Watson was disconnected from the Internet. Everything it knows about the world came from the corpus.
Average time to handle a question: three seconds.
2880 POWER7 cores (3.555GHz chip), four threads per core.
500GB per sec on-chip bandwidth (between the cores on a chip).
10Gb Ethernet network.
15TB of RAM.
20TB of disk, clustered. (Watson built its semantic Web from the 400TB corpus. It keeps the semantic Web, but not the corpus.)
Runs IBM DeepQA software, which has open-source components: Apache Hadoop distributed filesystem and Apache UIMA for natural language processing.
One full-time sysadmin on staff.
Ten compute racks, 80kW of power, 20 tons of cooling (for comparison, a human has one brain, which fits in a shoebox, can run on a tuna-fish sandwich and can be cooled with a handheld paper fan).
How Does Watson Work?
First, Watson develops a semantic net. Watson takes a large volume of text (the corpus) and parses that with natural language processing to create "syntatic frames" (subject→verb→object). It then uses syntactic frames to create "semantic frames", which have a degree of probability. Here's an example of semantic frames:
Inventors patent inventions (.8).
Fluid is a liquid (.6).
Liquid is a fluid (.5).
Why isn't the probability 1 in any of these examples? Because of phrases like "I speak English fluently". They tend to skew the numbers.
To answer questions, Watson uses Massively Parallel Probabilistic Evidence-Based Architecture. It uses the evidence from its semantic net to analyze the hypotheses it builds up to answer the question. You should watch the video of Michael's presentation and look at the slides, as there is really too much under the hood to present in a short article, but in a nutshell, Watson develops huge amounts of hypotheses (potential answers) and uses evidence from its semantic Web to assign probabilities to the answers to pick the most likely answer.
There are many algorithms at play in Watson. Watson even can learn from its mistakes and change its Jeopardy strategy.
Watson Is Built on Open Source
Watson is built on the Apache UIMA framework, uses Apache Hadoop, runs on Linux, and uses xCAT and Ganglia for configuration management and monitoring—all open-source tools.
Interview with Eddie Epstein on System Administration of the Watson Supercomputer
Eddie Epstein is the IBM researcher responsible for scaling out Watson's computation over thousands of compute cores in order to achieve the speed needed to be competitive in a live Jeopardy game. For the past seven years, Eddie managed the IBM team doing ongoing development of Apache UIMA. Eddie was kind enough to answer my questions about system administration of the Watson cluster.
AT: Why did you decide to use Linux?
EE: The project started with x86-based blades, and the researchers responsible for admin were very familiar with Linux.
AT: What configuration management tools did you use? How did you handle updating the Watson software on thousands of Linux servers?
EE: We had only hundreds of servers. The servers ranged from 4- to 32-core machines. We started with CSM to manage OS installs, then switched to xCat.
Aleksey Tsalolikhin has been a UNIX/Linux system administrator for 14 years.