Remote Direct Memory Access Networking for HPC: Comparative Review of 10GbE iWARP and InfiniBand

By Saqib Jang

February 24, 2010

The Rise of HPC Cluster Computing

While the HPC market is expected to experience a revenue dip in 2009, growth is expected to resume in 2010 and remain a bright spot in the overall IT market. The most important feature of the HPC growth trend is that it will continue to be fueled primarily by purchases of Linux cluster systems priced under $250,000. Cluster computing systems, separate compute nodes built from standard component technologies have caused disruptive changes in the HPC market.

As the component technologies of cluster systems have improved and buyers have become more confident running cluster systems, they have inevitably redirected capital once earmarked for large custom systems to larger cluster systems. These much larger clusters, often with thousands of processors, present opportunities for huge performance gains through improved parallel performance resulting in an overall higher order of magnitude return-on-investment (ROI). While algorithm and application tuning is often required to obtain these benefits, so often are cost, bandwidth, message rate, and latency of cluster interconnects.

One consequence of the range of requirements for cluster networking is that the leading interconnects in HPC are Gigabit Ethernet (which is based on Ethernet networking standard) and InfiniBand (delivering upwards of 10X performance vs. GbE). Both show significant deployment in HPC. The latest TOP500 list of HPC systems has 259 Gigabit Ethernet-based deployments compared to 181 InfiniBand-connected systems. The deployment of 10 Gigabit Ethernet (10GbE) cluster networking is emerging at this point. The price of this interconnect has been falling as the volume of its shipments grow. This growth is based on a combination of its 10X performance over GbE along with the ease of deployment due to its Ethernet heritage positions it for a bright future as a cluster interconnect.

As cluster systems have grown, so has the total amount of data in play in the average parallel HPC application. This has significant implications for HPC storage systems. Storage systems need to have the best possible bandwidth and latency characteristics. HPC storage systems have themselves become increasingly clustered and parallel as well as network-attached and accessible from all nodes on the cluster through the interconnect. In this context, the demand for interconnect solutions that supports a converged storage and cluster interconnect fabric is expected to grow significantly.

10GbE iWARP Overview and Value Proposition

For years, Ethernet has been the de facto standard LAN for connecting users to each other and to network resources. Ethernet sales volumes make it unquestionably the most cost-effective datacenter fabric to deploy and maintain. The latest generation of Ethernet, 10 Gigabit Ethernet (10GbE), offers a 10 Gbps data rate, which simplifies growth for existing data networking applications while removing the bandwidth barriers to deployment for highest-performance HPC clustering and storage networking.

  • 10GbE end-to-end performance now compares very favorably with that of more specialized datacenter interconnects, which eliminates performance as a drawback to the adoption of an Ethernet unified data center fabric.
     
  • Off-loading cluster and storage protocol processing from the central CPU to intelligent 10GbE NIC can also improve the power efficiency of end stations because off-load ASIC processors are generally considerably more power-efficient in executing protocol workloads.
     
  • The value of implementing TCP/IP protocol processing in silicon at 10 Gbps data rates is clear. Effectively, such approaches have the potential of reducing the relative bandwidth and latency overhead effect of TCP/IP protocol processing to zero.

Achieving 10GbE performance for latency-sensitive HPC communications has required solving Ethernet’s long-standing overhead problems; problems that, in slower Ethernet generations, were adequately overcome by steadily increasing CPU clock speeds.

Enter 10GbE iWARP

The iWARP extensions to TCP/IP focus on eliminating the three major sources of networking overhead — transport (TCP/IP) processing, intermediate buffer copies, and application context switches — that collectively account for nearly 100 percent of CPU overhead related to networking. Specifically, iWARP implements a number of mechanisms to provide a low-latency means of passing RDMA over Ethernet.

The iWARP extensions utilize advanced techniques to reduce CPU overhead, memory bandwidth utilization, and latency by a combination of offloading TCP/IP processing from the CPU, eliminating unnecessary buffering, and dramatically reducing expensive operating system calls and context switches — moving data management and network protocol processing to an accelerated RDMA over TCP/IP NIC (or R-NIC) 10 Gigabit Ethernet adapter.

R-NICs can reduce CPU utilization for 10 Gbps transfers to less than 10 percent and can reduce the host component of end-to-end latency to as little as 5–10 microseconds. High port-count 10GbE switches are available, which delivers HPC-class latency performance within 100’s of nanoseconds.

InfiniBand Overview and Value Proposition

InfiniBand is an I/O architecture designed to increase the communication speed between CPUs, devices within servers and subsystems located throughout a network. The original goal behind the release of the InfiniBand specification by the InfiniBand Trade Association was to address the mismatch between the speed of CPUs and the PCI I/O bus, as well as other deficiencies of the PCI bus, including bus sharing, scalability, and fault tolerance.

InfiniBand is a point-to-point, switched I/O fabric architecture. Both devices at each end of a link have full access to the communication path. To go beyond a point and traverse the network, switches come into play. By adding switches, multiple points can be interconnected to create a fabric. As more switches are added to a network, aggregated bandwidth of the fabric increases. By adding multiple paths between devices, switches also provide a greater level of redundancy.

A single InfiniBand link supports 2.5 Gbps in each direction per connection. InfiniBand supports double (DDR) and quad data rate (QDR) speeds, for 5 Gbps or 10 Gbps respectively, at the same data-clock rate. InfiniBand links use 8B/10B encoding — every 10 bits sent carry 8 bits of data, which meansthe net data transmission rate is four-fifths the raw rate. Thus single, double, and quad data rates carry 2, 4, or 8 Gbps respectively.

A quad-rate 12X link therefore carries 120 Gbps raw, or 96 Gbps of useful data. At present, most systems use 4X 10 Gbps (SDR), 20 Gbps (DDR) or 40 Gbps (QDR) connections. However, InfiniBand QDR performance is bounded by the 26 Gbps PCIe Gen2 throughput limitation.

Latency performance of InfiniBand SDR ad DDR switch chips is around 200 nanoseconds. InfiniBand Host Channel Adapters (HCAs) are rated 1-3 microseconds (although effective application-level performance is a different matter).

High-end clustering architectures have provided the main opportunity for InfiniBand deployment. Using the InfiniBand fabric versus Gigabit Ethernet as the cluster inter-process communications (IPC) interconnect typically boosts cluster performance and scalability while improving application response times. InfiniBand also provides exceptional scalability and failover in comparison to Gigabit Ethernet. In short, compared to Gigabit Ethernet, InfiniBand stands out in providing the mechanisms necessary to support the demanding requirements of high-end clustering.

iWARP and InfiniBand Comparative Review

As far as its compatibility with existing datacenter infrastructure, because it is layered on top of TCP, iWARP is fully compatible with existing Ethernet switching equipment that is able to process iWARP traffic out-of-the-box. In comparison, deploying InfiniBand requires environments where two separate network infrastructures are installed and managed as well as specialized InfiniBand to Ethernet gateways for bridging between the two infrastructures.

10GbE infrastructure is available from a range of incumbent and startup vendors. Intel, Broadcom, and Chelsio provide 10GbE iWARP adapters, while 10GbE switches are available from a broad range of vendors including Cisco, HP, IBM, BLADE Network Technologies, Extreme, Force10, Arista, and Voltaire. InfiniBand host channel adapter and switch silicon is only available from two vendors (Mellanox and QLogic), who in turn have signed up a number of OEMs to carry adapter and switching systems.

Both interconnects offer equivalent capabilities for supporting operating systems. The OpenFabrics software stack that is fully integrated into the flavors of Linux distributed by Novell and Red Hat fully supports both 10GbE iWARP and InfiniBand.

10GbE iWARP leverages its heritage to also support acceleration of emerging Ethernet-based storage protocols, including file storage (NFS-RDMA), which is fully supported by the Linux OFED stack. In addition, the Linux OFED stack also enables 10GbE iWARP to out-of-the-box support Lustre networking (LNET). In addition, 10GbE iWARP adapters can also provide concurrent, native support for standard Ethernet protocols such as NFS, CIFS, and iSCSI. In comparison, InfiniBand has had minimal deployments for server-to-storage communications, whether for file or block storage.

Regarding pricing, major server vendors are starting to add a 10 Gigabit Ethernet chip to the motherboard-known as LAN-on-Motherboard (LOM). NIC prices will continue to drop as LOM technology lets NIC vendors reach the high volumes they need to keep costs down, which in turn will drive switch port prices down as well. InfiniBand, on the other hand, has reached a mature market position and, consequently, reductions in the pricing of InfiniBand products will be relatively gradual.

Large-scale clusters built using 10GbE iWARP technology and high port-count 10Gbe switches are gaining ground, and cluster scalability is no longer viewed as inhibiting 10Gbe deployment. InfiniBand technology is an established interconnect for building large node-count clusters.

From a roadmap standpoint, the Ethernet market is moving forward aggressively to develop and implement 40G and 100G-based standards. It is expected that the standard for these versions of Ethernet will be ratified during 2010 and initial implementations based on these standards will be shipping from a range of vendors in the blade server and Ethernet networking switch markets within the next 2 to 3 years.

Converged Enhanced Ethernet

The IEEE has been developing standards collectively referred to as “Data Center Bridging” (DCB) or “Converged Enhanced Ethernet” (CEE) This refers to high speed Ethernet (currently 10 Gbps, with a clear path to 40 Gbps and 100 Gbps), plus a number of new features. The main new features are:

  • Priority-Based Flow Control (802.1Qbb), sometimes called “per-priority pause”
  • Enhanced Transmission Selection (802.1Qaz)
  • Congestion Notification (802.1Qau)

The first two features allow splitting an Ethernet link into multiple “virtual links” that operate independently — bandwidth can be reserved for a given virtual link, and by having per-virtual-link flow control, CEE can ensure that certain traffic classes do not overrun their buffers thus avoiding dropping packets. This congestion notification capability means that we can tell senders to slow down to avoid congestion spreading caused by that flow control.

CEE was developed primarily for use in Fibre Channel over Ethernet (FCoE). FC requires a very reliable network — it simply does not work if packets are dropped because of congestion — and, so, CEE provides the ability to segregate FCoE traffic on top of a “no drop” virtual link.

The roadmap initiatives in the InfiniBand space consist of QDR, EDR (2011), and RDMA over CEE. However, these roadmap initiatives suffer from the same limitations that have been a traditional challenge for InfiniBand, namely, limited vendor support.

RoCEE Overview and Value Proposition

Mellanox, the leader in the InfiniBand market, is behind the emerging RDMA over Converged Enhanced Ethernet (RoCEE) protocol proposal. RoCEE is designed to allow the deployment of RDMA semantics on Converged Enhanced Ethernet fabric by running the IB transport protocol using Ethernet frames.

Mellanox’s RoCEE proposal was motivated in order to create a protocol analogous to FCoE for Ethernet-based cluster networking. In other words, to take the InfiniBand transport layer and package it into Ethernet frames, instead of using the iWARP protocol for Ethernet-based high-performance cluster networking. But there are a number of challenges associated with this proposal:

First, one of the major motivations behind the RoCEE proposal is that it is the fastest path forward for an Ethernet-based alternative to InfiniBand. However, this ignores the fact that iWARP adapters are already shipping from multiple vendors, including Intel, Chelsio, and Broadcom. In addition, iWARP will automatically leverage the performance benefits of CEE as support for it will be ubiquitous in all 10GbE server adapter and LOM implementations, iWARP and non-iWARP alike.

Second, the idea that an InfiniBand over Ethernet (IBoE) specification will be quick or easy to develop flies in the face of the experience with FcoE. While FCoE sounded simple in concept, it turns out that the standards work took at least three years. In comparison, IBoE is more complicated to specify, and fewer resources are available for it, so a realistic view is that a true standard is very far away.

Last, RoCEE proponents point to the performance overhead challenges related to iWARP based on the TCP/IP protocol. However, this does not take into account the efficiency of silicon-based implementations of 10 Gbps TCP/IP. In addition, iWARP is also positioned to automatically take advantage of CEE as that protocol gains ubiquity in 10GbE server LOM and adapters.

In summary, RoCEE is unproven and its deployment faces significant hurdles including standardization and application and upper layer adoption. In addition, RoCEE is dependent on the deployment of 10GbE CEE infrastructure; currently only one vendor (Cisco) offers CEE switches, which are at relatively high price points.

About the Author

Saqib Jang is founder and principal at Margalla Communications, a Woodside, Calif.-based strategic and technical marketing consulting firm focused on storage and server networking.

This article is an excerpt from a Margalla Communications white paper entitled High-speed Remote Direct Memory Access (RDMA) Networking for HPC: Comparative Review of 10GbE iWARP and InfiniBand available at www.margallacomm.com.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

NSF Project Sets Up First Machine Learning Cyberinfrastructure – CHASE-CI

July 25, 2017

Earlier this month, the National Science Foundation issued a $1 million grant to Larry Smarr, director of Calit2, and a group of his colleagues to create a community infrastructure in support of machine learning research Read more…

By John Russell

DARPA Continues Investment in Post-Moore’s Technologies

July 24, 2017

The U.S. military long ago ceded dominance in electronics innovation to Silicon Valley, the DoD-backed powerhouse that has driven microelectronic generation for decades. With Moore's Law clearly running out of steam, the Read more…

By George Leopold

Graphcore Readies Launch of 16nm Colossus-IPU Chip

July 20, 2017

A second $30 million funding round for U.K. AI chip developer Graphcore sets up the company to go to market with its “intelligent processing unit” (IPU) in 2017 with scale-up production for enterprise datacenters and Read more…

By Tiffany Trader

HPE Extreme Performance Solutions

HPE Servers Deliver High Performance Remote Visualization

Whether generating seismic simulations, locating new productive oil reservoirs, or constructing complex models of the earth’s subsurface, energy, oil, and gas (EO&G) is a highly data-driven industry. Read more…

Trinity Supercomputer’s Haswell and KNL Partitions Are Merged

July 19, 2017

Trinity supercomputer’s two partitions – one based on Intel Xeon Haswell processors and the other on Xeon Phi Knights Landing – have been fully integrated are now available for use on classified work in the Nationa Read more…

By HPCwire Staff

NSF Project Sets Up First Machine Learning Cyberinfrastructure – CHASE-CI

July 25, 2017

Earlier this month, the National Science Foundation issued a $1 million grant to Larry Smarr, director of Calit2, and a group of his colleagues to create a comm Read more…

By John Russell

Graphcore Readies Launch of 16nm Colossus-IPU Chip

July 20, 2017

A second $30 million funding round for U.K. AI chip developer Graphcore sets up the company to go to market with its “intelligent processing unit” (IPU) in Read more…

By Tiffany Trader

Fujitsu Continues HPC, AI Push

July 19, 2017

Summer is well under way, but the so-called summertime slowdown, linked with hot temperatures and longer vacations, does not seem to have impacted Fujitsu's out Read more…

By Tiffany Trader

Researchers Use DNA to Store and Retrieve Digital Movie

July 18, 2017

From abacus to pencil and paper to semiconductor chips, the technology of computing has always been an ever-changing target. The human brain is probably the com Read more…

By John Russell

The Exascale FY18 Budget – The Next Step

July 17, 2017

On July 12, 2017, the U.S. federal budget for its Exascale Computing Initiative (ECI) took its next step forward. On that day, the full Appropriations Committee Read more…

By Alex R. Larzelere

Women in HPC Luncheon Shines Light on Female-Friendly Hiring Practices

July 13, 2017

The second annual Women in HPC luncheon was held on June 20, 2017, during the International Supercomputing Conference in Frankfurt, Germany. The luncheon provid Read more…

By Tiffany Trader

Satellite Advances, NSF Computation Power Rapid Mapping of Earth’s Surface

July 13, 2017

New satellite technologies have completely changed the game in mapping and geographical data gathering, reducing costs and placing a new emphasis on time series Read more…

By Ken Chiacchia and Tiffany Jolley

Intel Skylake: Xeon Goes from Chip to Platform

July 13, 2017

With yesterday’s New York unveiling of the new “Skylake” Xeon Scalable processors, Intel made multiple runs at multiple competitive threats and strategic Read more…

By Doug Black

Google Pulls Back the Covers on Its First Machine Learning Chip

April 6, 2017

This week Google released a report detailing the design and performance characteristics of the Tensor Processing Unit (TPU), its custom ASIC for the inference Read more…

By Tiffany Trader

Nvidia Responds to Google TPU Benchmarking

April 10, 2017

Nvidia highlights strengths of its newest GPU silicon in response to Google's report on the performance and energy advantages of its custom tensor processor. Read more…

By Tiffany Trader

Quantum Bits: D-Wave and VW; Google Quantum Lab; IBM Expands Access

March 21, 2017

For a technology that’s usually characterized as far off and in a distant galaxy, quantum computing has been steadily picking up steam. Just how close real-wo Read more…

By John Russell

HPC Compiler Company PathScale Seeks Life Raft

March 23, 2017

HPCwire has learned that HPC compiler company PathScale has fallen on difficult times and is asking the community for help or actively seeking a buyer for its a Read more…

By Tiffany Trader

Trump Budget Targets NIH, DOE, and EPA; No Mention of NSF

March 16, 2017

President Trump’s proposed U.S. fiscal 2018 budget issued today sharply cuts science spending while bolstering military spending as he promised during the cam Read more…

By John Russell

CPU-based Visualization Positions for Exascale Supercomputing

March 16, 2017

In this contributed perspective piece, Intel’s Jim Jeffers makes the case that CPU-based visualization is now widely adopted and as such is no longer a contrarian view, but is rather an exascale requirement. Read more…

By Jim Jeffers, Principal Engineer and Engineering Leader, Intel

Nvidia’s Mammoth Volta GPU Aims High for AI, HPC

May 10, 2017

At Nvidia's GPU Technology Conference (GTC17) in San Jose, Calif., this morning, CEO Jensen Huang announced the company's much-anticipated Volta architecture a Read more…

By Tiffany Trader

Facebook Open Sources Caffe2; Nvidia, Intel Rush to Optimize

April 18, 2017

From its F8 developer conference in San Jose, Calif., today, Facebook announced Caffe2, a new open-source, cross-platform framework for deep learning. Caffe2 is the successor to Caffe, the deep learning framework developed by Berkeley AI Research and community contributors. Read more…

By Tiffany Trader

Leading Solution Providers

How ‘Knights Mill’ Gets Its Deep Learning Flops

June 22, 2017

Intel, the subject of much speculation regarding the delayed, rewritten or potentially canceled “Aurora” contract (the Argonne Lab part of the CORAL “ Read more…

By Tiffany Trader

Reinders: “AVX-512 May Be a Hidden Gem” in Intel Xeon Scalable Processors

June 29, 2017

Imagine if we could use vector processing on something other than just floating point problems.  Today, GPUs and CPUs work tirelessly to accelerate algorithms Read more…

By James Reinders

Russian Researchers Claim First Quantum-Safe Blockchain

May 25, 2017

The Russian Quantum Center today announced it has overcome the threat of quantum cryptography by creating the first quantum-safe blockchain, securing cryptocurrencies like Bitcoin, along with classified government communications and other sensitive digital transfers. Read more…

By Doug Black

MIT Mathematician Spins Up 220,000-Core Google Compute Cluster

April 21, 2017

On Thursday, Google announced that MIT math professor and computational number theorist Andrew V. Sutherland had set a record for the largest Google Compute Engine (GCE) job. Sutherland ran the massive mathematics workload on 220,000 GCE cores using preemptible virtual machine instances. Read more…

By Tiffany Trader

Google Debuts TPU v2 and will Add to Google Cloud

May 25, 2017

Not long after stirring attention in the deep learning/AI community by revealing the details of its Tensor Processing Unit (TPU), Google last week announced the Read more…

By John Russell

Groq This: New AI Chips to Give GPUs a Run for Deep Learning Money

April 24, 2017

CPUs and GPUs, move over. Thanks to recent revelations surrounding Google’s new Tensor Processing Unit (TPU), the computing world appears to be on the cusp of Read more…

By Alex Woodie

Six Exascale PathForward Vendors Selected; DoE Providing $258M

June 15, 2017

The much-anticipated PathForward awards for hardware R&D in support of the Exascale Computing Project were announced today with six vendors selected – AMD Read more…

By John Russell

Top500 Results: Latest List Trends and What’s in Store

June 19, 2017

Greetings from Frankfurt and the 2017 International Supercomputing Conference where the latest Top500 list has just been revealed. Although there were no major Read more…

By Tiffany Trader

  • arrow
  • Click Here for More Headlines
  • arrow
Share This