David Fair, Unified Networking Mktg Mgr, Intel Networking Division
WARP was on display at IDF14 in multiple contexts. If you’re not familiar with iWARP, it is an enhancement to Ethernet based on an IETF standard that delivers Remote Direct Memory Access (RDMA). In a nutshell, RDMA allows an application to read or write a block of data from or to the memory space of another application that can be in another virtual machine or even a server on the other side of the planet. It delivers high bandwidth and low latency by bypassing the kernel of system software and avoiding interrupts and making extra copies of data. A secondary benefit of kernel bypass is reduced CPU utilization, which is particularly important in cloud deployments. More information about iWARP has recently been posted to Intel’s website if you’d like to dig deeper.
Intel is planning to incorporate iWARP technology in future server chipsets and systems-on-a-chip (SOCs). To emphasize our commitment and show how far along we are, Intel showed a demo using the RTL from that future chipset in FPGAs running Windows* Server 2012 SMB Direct and doing a boot and virtual machine migration over iWARP. Naturally it was slow – about 1 Gbps - since it was FPGA-based, but Intel demonstrated that our iWARP design is already very far along and robust. (That’s Julie Cummings, the engineer who built the demo, in the photo with me.)
Jim Pinkerton, Windows Server Architect, from Microsoft joined me in a Poster Chat on iWARP and Microsoft’s SMB Direct technology, which scans the network for RDMA-capable resources and uses RDMA pathways to automatically accelerate SMB-aware applications. No new software and no system configuration changes are required for system administrators to take advantage of iWARP.
Jim Pinkerton also co-taught the “Virtualizing the Network to Enable a Software Defined Infrastructure” session with Brian Johnson of Intel’s Networking Division. Jim presented specific iWARP performance results in that session that Microsoft has measured with SMB Direct.
Lastly, the NVMe (Non-Volatile Memory Express) community demonstrated “remote NVMe” made possible by iWARP. NVMe is a specification for efficient communication to non-volatile memory like flash over PCI Express. NVMe is many times faster than SATA or SAS, but like those technologies, targets local communication with storage devices. iWARP makes it possible to securely and efficiently access NVM across an Ethernet network. The demo showed remote access occurring with the same bandwidth (~550k IOPS) with a latency penalty of less than 10 µs.
Intel is supporting iWARP because it is layered on top of the TCP/IP industry standards. iWARP goes anywhere the internet goes and does it with all the benefits of TCP/IP, including reliable delivery and congestion management. iWARP works with all existing switches and routers and requires no special datacenter configurations to work. Intel believes the future is bright for iWARP.