As the flow of data generated on social media explodes, companies are striving to crunch data more efficiently so that they can respond to customer feedback fast.
For Salesforce.com ' which mines social media feeds for big customers like Dell, Cisco and Gatorade ' the task was becoming monumental.
The provider of software-as-a-service business applications has to analyze 500 million Twitter posts each day and match them against a database of more than 1.6 million search expressions. Over time, this created latency issues, as it takes about five minutes to process a batch of 8,000 tweets, a rate that's unacceptable for customers who want to respond in real time.
'Minutes are critical here,' Brendan Wood, a member of Salesforce's R&D staff, said during a presentation at the GPU Technology Conference. 'We're trying to shave off as much time as possible.'
Enter the GPU and CUDA. The company traded its old approach, which relied on Apache Lucene, the Java-based text search engine library, for one that featured CUDA-powered GPUs and an improved algorithm. The resulting approach, dubbed Zapp, relies on CPUs to handle keyword matching while GPUs handle the more compute-intensive tasks.
By building what is essentially a cost-effective, CUDA-powered search engine, Salesforce is now able to analyze the flow of incoming tweets with only two NVIDIA GTX 580 GPUs, freeing up hardware and personnel for other tasks, says Wood.
He estimates that CUDA helps Salesforce process searches 20-30 times faster than before for its customers. That's matters most to the companies that are trying to model responses to customer comments pulled from social media feeds.
Says Wood: 'Faster responses mean faster resolution of customer issues.'
And in the world of customer service, faster resolution means happier customers.