During the GTC 2014, the start-up SQream was awarded with the “company to watch” price for its innovative use of GPU in the database industry. More precisely in “Big Data” Analytics using generic SQL request. As usual with GPGPU technologies the idea is “simple”, its implementation often tricky and its business planning a kind of challenge. So our focus on their result which show explicit advantage compared to CPU cluster based solution.

SQream brought the entire database analytics workload on one GPU, with obvious advantage in price and performance. Today’s GPU allow around 6Gb to 12 Gb buffer plus memory streaming from host to device but still have bottleneck like bus bandwidth or no direct link between two node. SQream managed to handle the challenge of usual SQL processing on a dataset by internally tweaking data design and optimising it for GPU-specific advantage. Results showed up to be 100x faster running test queries on a simple GPU compared to a CPU based cluster. Though this sounds like the future of the sector, we don’t see any massive investment from major database player to the GPU technology, certainly due to their partnership with CPU based hardware and software maker.

^{Photo © 2014 EnterpriseTech. All Rights Reserved.}

Another problem or said challenge facing the SQream team is the lack of scalability of their solution due to communication constraint between cluster node. This problem is not specific to database object but to the strategy behind the implementation of a cluster. From child-board CPU based cluster solution (for example Intel Phi) to GPU based cluster and including all other more esoteric solution like Power-PC SPU based cluster, design must include the 2 scales segmenting the system: A first scale managing workload at a node (host) level, then a second scale handling workload at a (guest) level. So handling Petabyte of data using a SQream cluster will lead to a first segmentation of the database, each segment treated independently by its GPU. Future GPU architecture will fade out this constraint for sure, hopefully during the year 2015 with the Pascal technology announced by NVIDIA.