Infiniband in the Data Center

Infiniband in the Data Center Steven Carter Cisco Systems stevenca@cisco.com • Makia Minich, Nageswara Rao • Oak Ridge National Laboratory • {minich,rao}@ornl.gov

Agenda • Overview • The Good, The Bad, and The Ugly • IB LAN Case Study: Oak Ridge National Laboratory Center for Computational Sciences • IB WAN Case Study: Department of Energy’s UltraScience Network

Overview • Data movement requirements are, once again, exploding in the HPC community (sensors produce more data, larger computers compute with higher accuracy, disk subsystems are bigger/faster, etc) • The requirement to move 100’s of GB/s (the rates currently proposed for many of the new petascale systems) within the data center necessitate something more than is being currently provided by the Ethernet community • There also exists a requirement to move large amounts of data between data centers. TCP/IP does not adequately meet this need because of its poor wide-are characteristics. • This is a high-level overview of the pros and cons of using Infiniband to meet these needs and two case studies to reinforce them

Agenda • Overview • Infiniband: The Good, The Bad, and The Ugly • IB LAN Case Study: Oak Ridge National Laboratory Center for Computational Sciences • IB WAN Case Study: Department of Energy’s UltraScience Network

The Good • Cool Name (Marketing gets an A+ -- who doesn’t want infinite bandwidth?) • Unified Fabric/IO Virtualization: • Low-latency interconnect - nanoseconds, not low microseconds - not necessarily important in a data center • Storage – Using SRP (SCSI RDMA Protocol) or iSER (iSCSI Extension for RDMA) • IP – Using IPoIB, newer versions run over Connected Mode giving better throughput • Gateways – Gateways give access to legacy Ethernet (careful) and Fibre Channel networks

The Good (Cont.) • Faster link speeds: • 1x Single Data Rate (SDR) = 2.5 Gb/s (2 Gb/s with 8b/10b signalling) • 4 1x links can be aggregated into a single 4x link • 3 4x links can be aggregated into a single 12x link (single 12x link also available) • Double Data Rate (DDR) currently available, Quad Data Rate (QDR) on the horizon • Many link speeds available: 8Gb/s, 16Gb/s, 24 Gb/s, 32Gb/s, 48 Gb/s, etc.

The Good (Cont.) • HCA does much of the heavy lifting: • Much of the protocol is done on the Host Channel Adapter (HCA) heavily leveraging DMA • Remote Direct Memory Access (RDMA) gives the ability to transfer data between hosts with very little CPU overhead • RDMA capability is EXTREMELY important because it provides significantly greater capability from the same hardware

The Good (Cont.) • Nearly 10x less cost for similar bandwidth: • Because of its simplicity, IB switches cost less. Oddly enough, IB HCAs are more complex than 10G NICs, but are also less expensive. • Roughly $500 per port in the switch and $500 for a dual port DDR HCA • Because of RDMA, there is a cost savings in infrastructure as well (i.e. you can do more with fewer hosts) • Higher port density switches: • Switches available with 288 (or more) full-rate ports in a single chassis

The Bad • IB sounds too much like IP (Can quickly degrade into a “Who’s on first” routine) • IB is not well understood by networking folks • Lacks some of the features of Ethernet important in the Data Center: • Router – no way to natively connect two separate fabrics - The IB Subnet Manager (SM) is integral to the operation of the network (detects hosts, programs routes into the switch, etc). Without a router, you cannot have two different SMs for different operational or administrative domains (Can be worked around at the application layer). • Firewall – No way to dictate who talks to whom by protocol (partitions exist, but are too course grained) • Protocol Analyzers - They exist but are hard to come by, difficult to “roll your own” because of the protocol is embedded in the HCA

The Ugly • Cabling options: • Heavy gage cables with clunky CX4 connectors • Short distance (< 20 meters) • If mishandled, they have a propensity to fail • Heavy connectors can become disengaged • Electrical to optical converter • Long distance (up to 150 meters) • Uses multi-core ribbon fiber (hard to debug) • Expensive • Heavy connectors can become disengaged

The Ugly (Continued) • Cabling options: • Electrical to optical converter built on the cable • Long distance (up to 100 meters) • Uses multi-core ribbon fiber (hard to debug) • More cost effective than other solutions • Heavy connectors can become disengaged

Case Study: ORNL Center for Computational Sciences (CCS) • The Department of Energy established the Leadership Computing Facility at ORNL’s Center for Computational Sciences to field a 1PF supercomputer • The design chosen, the Cray XT series, includes an internal Lustre filesystem capable of sustaining reads and writes of 240GB/s • The problem with making the filesystem part of the machine is that it limits the flexibly of the Lustre filesystem and increases the complexity of the Cray • The problem with decoupling the filesystem from the machine is the high cost involved with to connect it via 10GE at the required speeds

CCS IB Network Roadmap Summary Ethernet core scaled to match wide-area connectivity and archive Infiniband core scaled to match central file system and data transfer Lustre Baker Gateway Ethernet [O(10GB/s)] Infiniband [O(100GB/s)] Jaguar High-Performance Storage System (HPSS) Viz

IB Switch Spider (Linux Cluster) Rizzo (XT3) XT3 LAN Testing • ORNL showed the first successful infiniband implementation on the XT3 • Using Infiniband in the XT3’s I/O nodes running a Lustre Router resulted in a > 50% improvement in performance and a significant decrease in CPU utilization

Observations • XT3's performance is good (better than 10GE) for RDMA • XT3's poor performance compared to the generic X86_64 host likely a result of PCI-X HCA (known to be sub-optimal) • In its role as a Lustre router, IB allows significantly better performance per I/O node allowing CCS to achieve the required throughput with fewer nodes than would be needed using 10GE

4x Infiniband SDR OC-192 SONET Host Host IB over WAN testing Ciena CD-CI (SNV) • Placed 2 x Obsidian Longbow devices between two test hosts • Provisioned loopback circuits of various lengths on the DOE UltraScience Network and ran test. • RDMA Test Results: Local: 7.5Gbps (Longbow to Longbow) ORNL <-> ORNL (0.2mile): 7.5Gbps ORNL <-> Chicago (1400miles): 7.46Gbps ORNL <-> Seattle (6600 miles): 7.23Gbps ORNL <-> Sunnyvale (8600 miles): 7.2Gbps DOE UltraScience Network Obsidian Longbow Obsidian Longbow Ciena CD-CI (ORNL)

Sunnyvale loopback (8600 miles) – RC problem

Observations • The Obsidian Longbows appear to be extending sufficient link-level credits • Native IB transports does not appear to suffer from the same wide-area shortcomings as TCP (i.e. Full rate with no tuning) • With the Arbel based HCAs, we saw problems: • RC only performs well at large messages sizes • There seems to be a maximum number of messages allowed in flight (~250) • RC performance does not increase rapidly enough even when message cap is not an issue • The problems seem to be fixed with the new Hermon-based HCAs…

Obsidian’s Results – Arbel vs. Hermon Arbel to Hermon Hermon to Arbel

Summary • Infiniband has the potential to make a great data center interconnect because it provides a unified fabric, faster link speeds, mature RDMA implementation, and lower cost • There does not appear to be the same intrinsic problem with IB in the wide-area as there is with IP/Ethernet, making IB a good candidate to transfer data between data centers

The End Questions? Comments? Criticisms? For more information: • Steven Carter • Cisco Systems • stevenca@cisco.com • Makia Minich, Nageswara Rao • Oak Ridge National Laboratory • {minich,rao}@ornl.gov

Infiniband in the Data Center