This article is more than 1 year old
Linux kernel-bypassing Quobyte plug-in goes with the TensorFlow for faster file access
Skip it, it's a Google thing
Linux-loving hyperscale types at Euro startup Quobyte have pushed out a plug-in for its Data Centre File System, used in HPC-style workloads, that enables TensorFlow apps to access its files directly instead of having to traipse through the Linux kernel.
BBC bypasses Linux kernel to make streaming videos flowREAD MORE
TensorFlow machine learning work can be read-intensive with lots of server cores and GPUs needing access to stored file data. Shaving time off each access by avoiding passing through the Linux kernel storage stack can shorten the time needed for model training and inference.
Kernel bypassing has often been used with applications bound by filesystem and network IO. For example:
- WekaIO uses Linux kernel bypass technology to speed its filesystem. It operates its own RTOS (real time operating system) in Linux user space, not kernel space, and runs its own scheduling and networking stack, with the latter talking directly to the host server's network interface card (NIC) through PCIe virtualisation. WekaIO also has its own memory management and talks directly to a server's SSDs.
- The BBC uses kernel bypassing for video-streaming to improve performance. It avoids video data being copied from the application's buffers to the socket buffer, and then from the socket buffer to the network device driver's buffers.
- Solarflare is in the kernel bypass game for its NVMe-over-TCP storage network fabric. Its XtremeScale NICs support a universal kernel bypass suite that accelerates NFV and storage server traffic. It says NVMe-over-TCP kernel extension provides its network adapters with the ability to approach the latencies of NVMe local flash storage over conventional POSIX-compliant Ethernet TCP networks with no RDMA required.
Quobyte claimed the tech could improve GPU utilisation, with less time wasted waiting for file IOs to complete. It can be used on-premises or in the Google cloud.
Other HPC parallel and scale-out file system suppliers may soon add TensorFlow support. The Reg storage desk sees Elastifile, Isilon, Lustre, Panasas, Qumulo, Spectrum Scale and WekaIO as likely candidates. ®