Oracle bypasses SAS/SATA controllers in flashy new servers Oracle has revealed two new servers using the NVM Express standard that does away with SAS/SATA controllers in ways the company says makes them perform very, very, well when stocked full of flash. The basics: the two new beasts are called the X5-2 and the X5-2L. Both are two-socket affairs and use Intel's Xeon E5-2600 v3 …
It's a 1U rackmount server.
Pizza boxes look like this: http://pizzasupplies.ca/images/oyster-pizza-box.png
which is why the term was used for systems like the Sun 3/80 http://www.sun3zoo.de/50/nvram-c.jpg
and SPARCstation 1/1+
1U rackmount servers don't look anything like pizza boxes.
Poor article and even poorer comments.
NVMe is an industry standard initiated by over 90 companies (so its not just Oracle). The first chip sets were available in 2012 (so no-one could have done this 5 years ago), and there was a re-badge in 2014 with 65 companies involved to NVM Express and it's not just PCI Express SSD - in fact it was explicitly defined to address the issues that high throughput flash experiences in PCI Express.
The already released hardware products are by Samsung and Intel : http://www.nvmexpress.org/products/. It is already supported in Windows, Linux, Solaris, BSD et al.
"The NVM Express, Inc. is directed by a thirteen-member board of directors selected by the promoter group, which includes Avago Technologies, Cisco, Dell, EMC, HGST, Intel, Micron, NetApp, Oracle, PMC, Samsung, SanDisk and Seagate."
It's poor journalism and commentry when people can't even be bothered to look at Wikipedia or find the relevant standards website.
NVMe isn't proprietary to any one vendor. It's an industry standard. An industry standard that provides a vendor neutral way to add PCIe connected SSD's. Servers, Jbods and motherboards are in process of coming to market from a number of different manufacturers. Mostly Server products. Google is your friend.
Most database systems today are I/O bound, not CPU bound. When these boxes have a max memory bandwith of 136GB/s, then if they're running a database or any IO intensive app then they need much more bandwith than PCIe SSDs can deliver to keep those CPUs busy. There's not much point in having most of the 36 cores just hanging around doing nothing while they wait for IO.
When we look at IO bound systems generally, then the bottleneck is getting the data from the disks/flash to the CPU fast enough. We have IOs galore on the back end and cheap CPUs with enormous capacity. Unfortunately most systems still connect with antiquated protocols that were originally designed for spinning disk and those connections will often be the bottleneck in overall system performance. NVMe has been designed specifically for the high performance gained from flash based storage.
...which is why you don't want to do physical (to disk, even SSD) I/O any more than absolutely necessary. Stuff it full of memory, reduce your buffer caches, and turn on the in-memory option with large buffers instead. Now your bottleneck is the RAM I/O speed due to SIMD instructions - and 12c isn't even using the latest version the CPUs have.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
