Making Watson Fast
Daniel Brown
• Need for Watson to be fast to play Jeopardy
– All computations have to be done in a few
– Initial application speed: 1-2 hours processing
time per question
• Unstructured Information Management
Architecture (UIMA): framework for NLP
applications; facilitates parallel processing
– UIMA-AS: Asynchronous Scaleout
• UIMA chosen at start for these reasons; other
optimization work only began after 2 years
(after QA accuracy/confidence improved)
UIMA implementation of DeepQA
UIMA implementation of DeepQA
• Type System
• Common Analysis Structure (CAS)
• Annotator
– CAS multiplier (CM): creates new “children” CASes
• Flow Controller
• CASes can be spread across multiple systems
(processed in parallel) for efficiency
Scaling out
• Two systems:
– Development (+question processing)
• Meant to analyze many questions accurately
– Production (+speed)
• Meant to answer one question quickly
Scaling out: UIMA-AS
• (UIMA-AS: Asynchronous Scaleout)
– Manages multithreading, communication between
processes necessary for parallel processing
• Feasibility test: simulated production system with
110 processes, 110 8-core machines
– Goal: less than 3 seconds; actual: more than 3
– Two sources of latency: CAS serialization, network
– Optimizing CAS serialization resulted in runtime of <1s
Scaling out: Deployment
• 400 processes, 72 machines
• How to find time
bottlenecks in such a
– Monitoring tool
– Integrated timing
measurements (in
flow controller
RAM Optimizations
• Wanted to avoid disk read/write time delays, so
all (production system) data was put into RAM
• Some optimizations:
Reference size reduction
Java object size reduction
Java object overhead
String size
Special hash tables
Java garbage collection with large heap sizes
• *Full GC between games
Indri Search Optimizations
• Indri search: used to find most relevant 1-2 sentences
from Watson database
• Using single processor, primary search takes too long
(i.e. 100s)
– Supporting evidence search even longer
• Solution?
– Divide corpus (body of information to search) into chunks,
then assign each search daemon a chunk
– (specifically, 50GB corpus of 6.8 million documents, 79
chunks of 100000 documents each, 79 Indri search
daemons with 8 CPU cores each; end result, 32 passage
queries could be run at once)
Preprocessing and Custom Content
• Watson must first analyze the passage texts
before being able to use them
– Deep NLP analysis - semantic/structural parsing,
• Since Watson had to be self-contained, this
analysis could be done before run time
– Used Hadoop (distributed file system software)
– 50 machines, 16GB/8 cores each
Preprocessing and Custom Content
• Retrieving the preprocessed data?
– Preprocessed data much larger than unprocessed
corpus (~300GB total)
– Built custom content server – allocated data to 14
machines, ~20GB each
– Documents then were accessed from these
End result
• Parallel processing combined with a number
of other performance optimizations resulted
in a final average latency of less than 3
– No one “silver bullet” solution

similar documents