• Cisco VIC with usNIC technology achieves 2.13 microseconds ping-pong latency using Open MPI across a Cisco Nexus 3548 Switch.
• Cisco VIC with usNIC technology achieves 1168 MBps of MPI ping-pong throughput.
• Cisco VIC with usNIC technology achieves 2336 MBps of MPI Exchange and MPI SendRecv throughput.
With the advent of highly dense multicore compute systems, the need for low-latency network communication between compute systems has become paramount. In addition, the cost of building such compute cluster systems, as well as the cost of managing them, has become a major consideration in the design and deployment of these systems.
This white paper presents a simple introduction to Cisco® user-space network interface card (usNIC) technology and also describes performance results using the Intel MPI Benchmark (IMB).
Cisco usNIC is a low-latency interface on top of Cisco UCS® Virtual Interface Card (VIC) 1225. The interface consists of a set of software libraries that enable a data path bypassing the kernel. Cisco UCS VIC 1225 is a converged network adapter (CNA) that supports Ethernet NICs and Fibre Channel host bus adapters (HBAs).
Cisco usNIC technology is a viable approach, since it enables ultra-low-latency network communication between nodes. As such, it can be deployed for various latency-dependent tasks, high-performance computing (HPC) clusters being one of them.
System under Test Configuration
System Hardware and Software
Hardware: Cisco UCS VIC 1225 CNA
Firmware: 2.1 (2.127)
Driver software: enic-188.8.131.52
Hardware: Cisco UCS C220 M3 Rack Server with Intel® E5-2690 processor with 1600-MHz DDR3 memory
The following BIOS settings were used for this performance testing:
Intel Hyper-Threading technology: Disabled
Number of enabled cores: All
Execute disable: Disabled
Intel VT: Enabled
Intel VT-d: Enabled
Intel VT-d coherency support: Enabled
Intel VT-d ATS support: Enabled
CPU performance: HPC
Hardware prefetcher: Enabled
Adjacent cache line prefetcher: Enabled
DCU streamer prefetch: Enabled
DCU IP prefetcher: Enabled
Power technology: Custom
Enhanced Intel SpeedStep® technology: Enabled
Intel Turbo Boost technology: Enabled
Processor power state C6: Disabled
Processor power state C1 enhanced: Disabled
Frequency floor override: Disabled
P-STATE coordination: HW_ALL
Energy performance: Performance
Select memory RAS: Maximum performance
DRAM clock throttling: Performance
Low-voltage DDR mode: Performance mode
DRAM refresh rate: 2x
Channel interleaving: Auto
Rank interleaving: Auto
Patrol scrub: Disabled
Demand scrub: Enabled
Other settings were left at default values.
Operating System Settings
The following kernel parameter enables support for the Intel I/O memory management unit (IOMMU): intel_iommu=on
The following parameter turns off the Intel CPU idle driver: intel_idle.max_cstate=0
The following kernel parameter explicitly disables the CPU C1 and C1E states: idle=poll
Note: Using the idle=poll kernel parameter can result in increased power consumption. Use with caution.
The above kernel parameters were configured in the/etc/grub.conf file.
The OS CPU governor was configured for "performance." In the file/etc/sysconfig/cpuspeed, the "GOVERNOR" variable was set to "performance."
The test nodes were also configured to operate in runlevel 3 to avoid unnecessary background processes.
In the file /etc/inittab, the following line replaced the OS default:
SELinux was disabled.
In the file/etc/selinux/config, the following line replaced the OS default:
The following vNIC adapter configuration was used:
Number of VFs instances = 16
Interrupt coalescing timer = 0
The vNIC is directly configured from the onboard Cisco Integrated Management Controller (IMC).
Network Switch Configuration
The Cisco Nexus 3548 was configured with the following settings:
Flow control: On
No Drop mode: Enabled
Network MTU: 9216
WARP mode: Enabled
To enable flow control on the Nexus 3548:
interface ethernet 1/1-48
flowcontrol receive on
flowcontrol send on
To enable No Drop mode and Pause:
class-map type network-qos class1
match qos-group 1
policy-map type network-qos my_network_policy
class type network-qos class1
service-policy type network-qos my_network_policy
show running ipqos
class-map type qos class1
match cos 2
policy-map type qos my_qos_policy
class type qos class1
set qos-group 1
service-policy type qos input my_qos_policy
To enable MTU 9216:
policy-map type network-qos jumbo
class type network-qos class-default
service-policy type network-qos jumbo
To enable WARP mode:
hardware profile forwarding-mode warp
copy running-config startup-config
Note 1: The above configuration is specific to the system under test presented here and may not be directly applicable to all use cases. Please consult your local network administrator or refer to the Cisco Nexus 3548 Command Reference (see the link in the appendix) for more information.
Note 2: The above configuration also specifies send/recv flow control and No Drop mode enabled. This prevents the switch from dropping packets on the network with a combined use of port buffer management and network pause. These settings are specific to the test. In some application instances, it may be optimal not to enable send/recv flow control and `no drop' mode.
Two nodes were connected to a single Cisco Nexus 3548 Switch. The Cisco Nexus 3548 is an ultra-low-latency-capable switch from Cisco that is well suited for low-latency network messaging. For more details, refer to the product page link for the switch in the appendix.
Figure 1 shows the network topology that was used.
Figure 1. Network Topology
NetPIPE version 3.7.1 was used for testing point-to-point latency and for a throughput comparison between Cisco usNIC and Kernel TCP.
Intel MPI Benchmarks (IMB) version 3.2.4 was used for testing. Refer to the links in the appendix for more information about this software.
NetPIPE performs a simple ping-pong test between two nodes and reports half-round-trip (HRT) latencies in microseconds and throughput in MBps for a range of message sizes. Figure 2 shows a graph of the test results.
The following mpi command was used to run the NetPIPE test:
hwloc-bind socket:0.core:0 \
--host n1,n2 \
--mca btl usnic, self, sm \
Figure 2. NetPIPE Latency and Throughput
Intel MPI Benchmarks
IMB runs a set of MPI tests between two nodes and reports latencies (HRT) and throughput in MBps for a range of messages at sizes between and including 2^0 and 2^22.
The following tests were run:
Allreduce, Reduce, Reduce_scatter
Gather, Gatherv, Scatter, Scatterv
Allgather, Allgatherv, Alltoall, Alltoallv, Bcast
For more information on the benchmarks, refer to the IMB user guide link in the appendix.
The following mpi command was used to run the IMB tests:
hwloc-bind socket:0.core:0 \
--host n1,n2 \
--mca btl usnic,self,sm \
/opt/IMB-MPI1 -iter 10000, 10000, 10000
Figures 3, 4, and 5 present IMB ping-pong latencies and throughput for a range of message sizes using Cisco usNIC. The message sizes are split up into small, medium, and large to allow a closer look at the trend.
Figure 3. IMB PingPong Small Message
Figure 4. IMB PingPong Medium Message
Figure 5. IMB PingPong Large Message
The rest of the tables present consolidated performance information from various IMB performance tests.
With a small packet MPI ping-pong latency of 2.13 microsec and a maximum ping-pong throughput of 1168 MBps the Cisco usNIC on the Cisco UCS VIC 1225 with the Cisco Nexus 3548 enables a full HPC stack solution that is capable of both low latency and high throughput. Therefore, it is a compelling approach to running HPC tasks with Open MPI on standard Ethernet networks.