Many data center operators often focus on the CPU power and storage for their new data center hardware, but it is also very important to give consideration to the Network Interface Card (NIC) which is an often overlooked element of data center hardware.
For data-intensive applications, the NIC can be a crucial component that affects overall system performance. In today’s complex networked environment, this seemingly simple part can no longer be thought of as consisting of commodity components to support critical workloads.
The Hidden Performance Bottleneck
When the ASUS Eee Top AS10 first arrived at bit-tech back in 2009, it featured support for three 1 Gigabit Ethernet ports – at the time, we felt then that they might not be sufficient for future networking requirements. Whilst 125 megabytes per second sounds quick, our tests revealed the BlurMotion video playback only pushed around 20mbps through the PC’s pipe.
When a database server is running at full capacity serving thousands of concurrent queries, every transaction generates byte streams of data that need to be sent over the network cord. Typical NICs handle these byte streams by placing them in a data queue on the NIC which can cause delays if many other processes are trying to access the same bandwidth at the same time.
We are currently experiencing slow system performance where applications may open slower than normal or possibly time out. Additionally systems may seem to lock up or seize up allowing no activity to occur.
The maths against running modern applications on older servers is stark. The explosion in data generated by applications such as video streaming, real-time big data analytics, machine learning and high-frequency trading far outstrips the bandwidth of ageing network interfaces in legacy servers. While IT managers might be fond of their existing hardware, the issue with running modern workloads is that they produce so much more data, and that data has to travel across the network much faster, meaning older servers are unable to cope. Everything else becomes irrelevant when the network card is the bottleneck.
How High-speed NICs Transform Throughput
Today’s networks operate at higher speeds with the help of High-Speed Network Interface Cards that support up to 400 Gigabit (Gb/s) i.e. 10 Gb/s, 25 Gb/s, 40 Gb/s, 100 Gb/s etc. However, the increase in speed is not linear.
Thanks to a 100 Gigabits per second connection, the NIC can process up to 12.5 gigabytes per second, which is 100 times faster than current Gigabit Ethernet networking. But it’s only in an environment that has multiple applications and protocols sharing a single network pipeline where such a massive leap really comes into play.
Due to their ability to manage multiple interfaces efficiently, high-speed cards process packets faster and handle more interfaces than low-speed cards. This improves performance and eliminates “chokepoints,” where slower cards would slow down other functions of your workflow.
Latency Reduction Matters More Than You Think
While throughput gets a lot of press, decreasing latency has a much greater impact on many programs.
For systems that provide a trading platform, control system or interactive multimedia application, the critical parameter is not the bandwidth, but the all important latency. High-end NICs can include specialised hardware to reduce the overall latency, often completely independent of the normal software drivers. In addition to low latency, the hardware can also support advanced queueing mechanisms to optimise real-time transmission. Some cards can even bypass the Linux kernel and deliver the packets directly to the application.
The AMD Pollara 400 network interface card is designed to reach microsecond latency while supporting 400 Gigabit connectivity, a feat today’s standard networking hardware cannot deliver.
CPU Offloading Frees Processing Power
Traditional Gigabit NICs increase the speed at which data is sent and received. The new breed of NICs, however, also take some of that network processing load off the CPU.
Traditional Network Cards relied on the Systems Processor to Inspect, Encrypt, and Direct packets. Newer designs incorporate their own Processors and run their own code to Inspect, Encrypt, and Direct packets. This shifts the workload from the Systems Processor to the Network Processor freeing the CPU to perform application specific functions where it is far more valuable.
Enabling CPU offloading within compute-intensive environments can increase system performance by up to 20-30% helping to direct more resources to applications that generate revenue.
Real-world Impact on Data Center Operations
Performance improvements translate into operational gains that matter. Employees spend less time running backup jobs, synchronizing databases, and performing VM migrations – all designed to minimize windows of downtime as much as possible.
Another key benefit is lower power consumption since high-speed NICs require less power per unit of bandwidth than do multiple lower-speed NICs. This allows a Data Center to be smaller but more efficient, delivering greater throughput while using less square footage.
Fewer cards reduce the chance of failure points, simplify management, and reduce power usage. The cost savings of running a faster network translate to serving more customers with your current data center infrastructure, reduced server requirements to achieve faster application performance, and a faster backup and disaster recovery process to minimize risk.
Implementation Requires Careful Planning
As people move to NICs with increased speed (e.g. GigE (1,000 mbps) or 10 GigE (10,000 mbps)), they need to remember that the switches and cables on the network also will need to be upgraded. Also many current applications will need to be tweaked to more efficiently take advantage of the increased speed.
These improvements pay for themselves within a few months. For data centre’s that support mission critical workloads, high-speed network infrastructure is not a “nice to have” but a necessity. Improved throughput and reduced network congestion allows for increased productivity to ensure that business running on these platforms are not impaired by basic infrastructure.
While high-speed Networking Infrastructure for the data center has been gaining attention in recent years, many are unaware of the performance that is left on the table by not utilizing faster NICs.
