LSI AND REDHAT were showing off the next step in hardware virtualization, IOV or I/O Virtualization, during IDF. The idea is simple, make SAS or SATA cards aware of virtual spaces, speeding up I/O for VMs many times.
The PCIe bus is not virtualization aware by default, it needs to support the IOV spec to play in virtual space. In this case, the sub-spec LSI used is called SR-IOV or Single Root IOV because there is only one physical PC that initiates the calls to storage. Instead of having the CPU or chipset do the virtual to physical, or physical to virtual address translation, IOV aware systems do it automatically in hardware. This saves hundreds of cycles for each I/O transaction.
SR-IOV mapping, 1:1 is better than many:1
IOV supporting systems and controllers will see either 16 or 32 virtual controllers, each of which appears to be a full card with interrupt support and all the functionality you would expect from a non-virtualized part. This means the system can do fully transparent wire speed I/O transfers without any configuration. This is all due to the 1:1 mapping between the VMs hardware accelerated virtual HBAs, something not possible before IOV.
The LSI demo at IDF was based on RedHat’s KVM hypervisor and an Intel 5500 system that supports SR-IOV. The part that mattered was a prototype LSI MegaRAID controller that was IOV aware. It is still in development, but you can expect to see this, or something very similar to it on the shelves in a bit over a year.
How well does it do? There were two identical test systems based on a Xeon 55xx server with full SR-IOV support. It was running RedHat’s KVM and had a MegaRAID card with several Intel SSDs attached for big numbers. One system had the IOV functionality turned on, the other had it turned off.
LSI thankfully does not support Windows DRM infections
When IOV was turned on, the system scored almost 32K IOPS in IOMeter with an average response time of .52ms and 41% CPU utilization. With it off, the numbers were a little down, not quite 16K IOPS with a response time of .96ms and 42% CPU utilization. Half the throughput with double the latency is not a good thing, even if the CPU time didn’t go through the roof.
As you can see, LSI is at the forefront of I/O virtualization. This demo of SR-IOV is the first that we know of, and while it is still a year or more from production, it already almost doubles I/O performance for VMs. Not a bad start.S|A
Latest posts by Charlie Demerjian (see all)
- Thing go bump(gate) in the night for Nvidia’s GP100 Pascal GPU - May 3, 2016
- Cavium’s Octeon TX blends compute and packet moving - May 2, 2016
- Qualcomm releases Zeroth API to developers - May 2, 2016
- Another detail about Qualcomm server SoCs revealed - Apr 27, 2016
- AMD finally really honestly launches the dual Fury - Apr 26, 2016