High Performance Interconnects: Landscape, Assessments & Rankings Dan Olds Partner, OrionX April 12, 2017
Specialized 100G InfiniBand MPI Multi-Rack OPA Single-Rack HPI market segment 100G 40G 10G TCP/IP JDBC RMI IIOP SOAP etc. Current HPI line 1G Link Speed Net Protocol App Comm
2016 OrionX 4 Three Types of HPI Ethernet Sold by a host of providers, Cisco, HPE, Juniper, plus many others Tried and true interconnect, easiest to implement While it has the bandwidth of others, latency is pretty high (ms rather than ns) Proprietary Primarily sold by Cray, SGI, and IBM plus a few others Have to purchase a system in order to get their brand of HPI Intel is a new entrant in this segment of the market, although without an accompanying system InfiniBand Mellanox has emerged as the de-facto leader Highest performance based on published numbers:200gb/s, 200m messages/s, 90ns latency
Key Differences in HPI: Product Maturity/Position Ethernet Ethernet has been around longer than any HPI, but was surpassed in performance Still many installations, but has lost much of its share at the high end Latency (measured in ms, not ns) is the problem, not bandwidth 2016 OrionX 5
2016 OrionX.net 6 Key Differences in HPI: Product Maturity/Position Intel Omni Path Architecture Intel and Omni-Path are in their infancy still, very few installations Handful of customers (although some big names), few, if any, in production Claims bandwidth/latency/message rate same or better than InfiniBand (covered later)
Key Differences in HPI: Product Maturity/Position InfiniBand Has been in the HPI market since early 2000 s Thousands of customers, millions of nodes Now makes up large proportion of Top500 list (187 systems) Synonymous with Mellanox these days 2016 OrionX.net 7
2016 OrionX 8 Key Differences in HPI technology Onload vs. Offload Onload: main CPU handles all network processing chores, adapter and switches just pass the messages, examples Intel Omni-Path Architecture, Ethernet Also PC servers, old UNIX systems where CPUs handled every task and received interrupts on communications Offload: HCA and switches handle all network processing tasks, very little or no need for main CPU cycles, allows CPU to continue processing applications, examples Mellanox InfiniBand Mainframes with communication assist processors used to allow CPU to process applications, not communications
2016 OrionX.net 9 Offload Details Network protocol load includes: Link Layer: packet layout, packet forwarding, flow control, data integrity, QoS Network layer: adds header, routing of packets from one subnet to another Transport layer: in-order packet delivery, divides data into packets, receiver reassembles packets, sends/receives acknowledgements MPI operations: scatter, gather, broadcast, etc. With offload, ALL of these operations are handled by the adapter hardware, example: InfiniBand HBA
2016 OrionX.net 10 Onload Details Network protocol load includes: Link layer packet layout, packet forwarding, flow control, data integrity, QoS Network layer: header, routing of packets from one subnet to another Transport layer: in-order packet delivery, divides data into packets, receiver reassembles packets, sends/receives acknowledgements MPI operations: scatter, gather, broadcast, etc. With onload, ALL of these operations are performed by the host processor, using host memory
2016 OrionX.net 11 Onload vs. Offload Onload vs. Offload isn t a big deal when the cluster is small
2016 OrionX.net 12 Onload vs. Offload But it will become a very large deal when the cluster becomes larger Will particularly be a problem on scatter, gather type collective problems when head node will be overrun trying to process messages
Onload vs. Offload As node count increases, performance of Onload will drop Higher node count = more messaging, pressure on head node Node counts are increasing significantly Offload uses dedicated hardware ASICs Much faster than general purpose CPUs MPI not highly parallel With Onload, this means that speed is limited to slowest core speed Has no bearing on Offload speed 2016 OrionX.net 13
Rampant FUD War From: The Next Platform, Intel Stretches Deep Learning On Scalable System Framework, 5/10/16 Cost of HPI in cluster budget is typically ~15-20% of total Prices in high tech typically don t increase over time Price points for new products typically are the same as the former high end product they replace ex: high end PCs, low-end servers, etc. 2016 OrionX.net 14
2016 OrionX.net 15 More FUD War.. All images provided by Intel, all from The Next Platform story Intel Stretches Learning on Scalable System Framework May 10 th, 2016 What else do these images have in common?
2016 OrionX.net 16 FUD Wars Behind the Numbers It s all in the fine print, right? Here s Intel s fine print for the graphs on the last slide. dapl is key, it s an Intel MPI mechanism that doesn t allow for offload operations ala InfiniBand..48 port (B0 silicon). IOU Non-posted Prefetch disabled in BIOS. Snoop hold-off timer = 9. EDR based on internal testing: Intel MPI 5.1.3, shm:dapl fabric, RHEL 7.2 -genv I_MPI_DAPL_EAGER_MESSAGE_AGGREGATION off. Mellanox EDR ConnectX-4 Single Port Rev 3 MCX455A HCA. Mellanox SB7700 36 Port EDR InfiniBand switch. MLNX_OFED_LINUX-3.2-2.0.0.0 (OFED-3.2-2.0.0). IOU Non-posted Prefetch enabled in BIOS. 1. osu_latency 8 B message. 2. osu_bw 1 MB message. 3. osu_mbw_mr, 8 B Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.
Even More FUD 100% CPU core utilization on a Offload HCA?!! Does anyone believe this?!! This means that about half of the Top500 systems are absolutely useless Intel is using a CPU polling mechanism that pegs the CPU on the Mellanox box at 100%, yet has nothing to do with network comms Both Intel and Mellanox have benchmarked OPA at ~65% CPU utilization 2016 OrionX.net 17
2016 OrionX.net 18 FUD Aside, here are the numbers.update FOR HDR Intel OPA Mellanox EDR/HDR InfiniBand Bandwidth 100 Gb/sec 100 Gb/sec / 200Gb/sec Latency (ns).93.85 or less /.90 or less Message rate 89 million/sec* 150m/sec / 200m/sec * this number, provided by Intel, has dropped from >150 million in 2015
2016 OrionX.net 19 HPI Roadmaps InfiniBand roadmap shows HDR now (200Gb/ s) and NDR down the road (400Gb/s? 2020?) OPA Roadmap Formerly: OPA 2 in 2018 Now: OPA 2 in 2020 ouch Ethernet roadmap shows 200Gb/s in 2018-19
2016 OrionX.net 20 Major HPI Choices: OrionX analysis Vendor Market Customer Product Presence Trends Overall Readiness Needs Overall Capabilities Roadmap Overall Mellanox 9 9 9 8 9 8.5 9 10 9.5 Ethernet vendors 7 7 7 9 6 7.5 7 6 6.5 Intel 6 8 7 6 7 6.5 7 7 7.0
OrionX Constellation Mellanox Intel Ethernet vendors Vendor Market Product Customer Ethernet 7 6.5 7.5 Mellanox 9 9.5 8.5 Intel 7 7.25 6.5
2016 OrionX 22 Questions? Comments? Concerns? OrionX Constellation reports