Fujitsu hits the data accelerator and parallelises SSD access
Hopes to bring this into practical use during mid 2017
Fujitsu Labs has had in-memory database software directly talk to an SSD, optimising data positioning for faster access, and boosting database performance threefold over standard SSD use.
The in-memory database SW sent read and write commands directly to the flash chips. It has a read-ahead feature, cognisant of data access patterns, that enabled parallel data retrievals from multiple flash memory chips.
“By loading data into DRAM before using the data from the in-memory database, data usage and data loading are processed simultaneously," said a Fujitsu release, "enabling high-speed big data processing, with no access lags, even with limited DRAM capacity."
Normally, system software issues read/write requests to an SSD in series and some of these requests can target the same flash chip, causing a wait. If the desired data was spread across several flash chips then they could be read in parallel, speeding access.
This is what the Fuji Labs SW does, with its software directly addressing the flash chips and distributing read instructions from an in-memory database among multiple flash memory chips to operate in parallel.
[Its SW] analyses the memory access patterns of an in-memory database and, instead of retrieving the data from flash memory each time, retrieves the data in parallel from multiple flash memory chips in advance so that the data are read instantaneously, without competing for access.
It developed a SW-controlled PCIe SSD, with 16 control channels and 256 flash chips*. This delivered 5.5GB/sec of bandwidth.
Fujitsu Labs aims to increase data retrieval parallelisation and check out applications beyond in-memory databases. It hopes to bring this technology into practical use in its fiscal 2017 year, which starts 1 April, 2017. ®
* Fujitsu's statement categorically said 256 flash chips, not 256Gbit flash chips.