December 2, 2013

GPUs Ease Network Monitoring Chore

Tiffany Trader
G-NetMon Figure1

Accelerators, like NVIDIA Tesla GPUs and Intel Xeon Phi coprocessors, made a big splash this year at SC13 as these highly parallel number-crunchers carved out a significant presence on both the TOP500 and Green500 charts, but there’s another promising use for accelerators, according to Fermilab researcher Wenji Wu: network monitoring.

Traditional network analysis tools have struggled to keep pace with the traffic demands as network bandwidth has skyrocketed. Adding to the strain, network administrators expect to be able to examine packets in real-time.

Wu, a network researcher at the U.S. Department of Energy’s Fermi National Accelerator Laboratory, believes that GPUs may offer some advantages over the current technology, which employs CPUs and ASICs. He delivered a paper presentation at SC13 last month, describing his research using off-the-shelf NVIDIA GPUs as network monitors in high-speed networks. According to Wu and his team, graphics chips are well-suited to the task.

Wu affirms that GPUs have “a great parallel execution model.” They offer high compute power, ample memory bandwidth, easy programmability, and can divvy up the processing duties into parallel tasks.

The Fermilab team, under Wu’s direction, built a prototype GPU-accelerated network performance monitoring system called G-NetMon to support large-scale scientific collaborations. The G-NetMon system consists of two 8-core 2.4 Ghz AMD Opteron 6136 processors, two Gbps Ethernet interfaces, 32 GB of system memory and one Tesla C2070 Fermi GPU.

“Our system exploits the data parallelism that exists within network flow data to provide fast analysis of bulk data movement between Fermilab and collaboration sites,” notes Wu. “Experiments demonstrate that our G-NetMon can rapidly detect sub-optimal bulk data movements.”

G-NetMon was designed to handle current network loads and also be able to accommodate expected future traffic demands. An experiment showed that the GPU-based prototype was between nine to 17 times faster than a single-core CPU. When compared to a six-core CPU, the GPU setup was about 1.5 times to 3 times faster. The next step for the researchers is to add some security features.