DB dev waggles GPU-dosed POWER9 server at data warehouse crowd, yells: SQream
NVLinking CPUs to GPUs chews through workloads
GPU database-botherer SQream has said its DB runs up to 150 per cent faster when it uses IBM's POWER9 CPUs linked to the GPUs rather than x86 processors.
SQream's big idea is to run repetitive data warehouse routines on GPUs rather than on limited-in-number CPU cores and so bring GPU acceleration benefits from AI and machine learning to the staid old data warehouse world.
The firm has worked with Big Blue to support IBM's POWER9 server, understood to be the AC922, with attached Nvidia GPUs. Unlike x86 servers linked to Nvidia GPUs, the IBM system has Nvidia's NVLink interconnecting both the GPUs and the CPUs. This made the POWER9 server up to 150 per cent faster than x86 servers equipped with the same GPUs.
Putting GPUs in the database driver's seat
In an earlier comparison test (PDF) run by a network operator, SQream compared a traditional massively parallel processor system (MPP) chewing through data warehouse data to an X86 HPE ProLiant fitted with a GPU card.
In the test, the operator's existing MPP data warehouse, consisting of 40 compute nodes in five full racks, was compared to an HPE DL380 Gen 9 server with an Nvidia Tesla card. The test involved analysing a few months' worth of data, coming in at 1.6TB per week, and encompassing CDR (call data records) and non-CDR data, such as customer profiles and customer-registered products.
One test run had a 10-step report generated, querying data from 13 different tables and combining them into a single result-table used to identify top usage location, segmented by customer. The report used advanced SQL features, including window functions.
Table from SQream white paper
The MPP system took 2.48 hours to run this query while the dull old ProLiant, with its Tesla GPU, ran it 18 times faster, completing it in eight minutes and 16 seconds.
That's what happens when GPU cores chew through data instead of x86 cores.
Changing the x86 cores for Big Blue's POWER9, it would seem, has them chomping through the data even faster.
It's worth noting that the whole GPU-drive database area is new and none of the mainstream data warehouse suppliers have products in the area yet.
SQream was founded in 2010 and its total funding is $26.4m, with a $19m round earlier this year - cash used to develop its business infrastructure and product tech.
The startup will discuss the details in Amsterdam on 4 October at the OpenPOWER Foundation Summit in a a session entitled "Tackling the Challenges of Massive Data Analytics". ®