Benefitting from AI-Enabled Mitigation of WAN Latency

Like many things, artificial intelligence has two sides to its coin: it can be good, and it can be bad. For example, it can be used by bad actors to hack and steal data and, on the plus side, it can be used to intelligently protect and automate networks and systems. Another benefit, thanks to WAN Acceleration, which deploys AI, machine learning and data parallelisation, is the ability to mitigate the effects of latency and packet loss. This means that encrypted data can be stored far apart and retrieved without consequence.

Without WAN Acceleration, the impact of latency and packet loss on data can be disastrous – meaning that it could take hours, days, or weeks more to back up and retrieve data, ensure business and service continuity ahead of any potential manmade or technological failure. Not being prepared could mean a loss of ability to analyse data, gain prompt access to customer records, or even to use cloud applications or e-commerce systems to deliver products and services.

Slower software-as-a-service (SaaS) and cloud apps could also lead to degraded collaboration, customer and partner interactions. Jitter caused by latency and packet loss or slow systems could increase user and customer frustration, as well as create higher numbers of IT helpdesk tickets. Then there is the growing need to maintain – particularly in industries such as financial services – regulatory compliance and oversight to prevent data loss and to avoid, for example, GDPR fines.

When data movement is slow, teams must structure their routines around it, setting transfers to run overnight or during downtime. This does more than reduce productivity; it limits our ability to understand and respond to the insights that data provides – slowing critical decision-making and making an organisation less able to act with strategic nimbleness. This, too, can be costly.

Traditional approaches

The trouble is that traditional fixes to improving WAN performance often fall on their backs. Adding more bandwidth and manual tuning doesn’t necessarily mean less latency and more speed. Elevated levels of latency can cripple TCP/IP transfer rates, even over a theoretically high-bandwidth WAN infrastructure: This is reason investing in new and increasingly sophisticated and expensive infrastructure doesn’t often provide the right solution to improving WAN performance.

WAN Optimisation has often been used to tackle latency and packet loss, but its biggest weakness is that it can’t handle encrypted data; even SD-WANs benefit from a boost with a WAN Acceleration overlay. It can also free up IT resources to focus on more strategic activities by automatically analysing data flows to minimise the impact of latency and packet loss, while boosting bandwidth by up to 98%.

Proof of concept evaluation

Teaming up with IBM Partner and Reseller, OCF, Bridgeworks conducted a proof-of-concept evaluation on IBM Storage Scale version 5.2.3.2. Also involved on the software end was IOR 4.0.0, IO500 SC20 benchmark suite, and VMware ESXi 8.0 Update 3. The hardware used was 2 Dell PowerEdge R730 as ESXi hosts, 2 PORTrockIT 400 Nodes (hardware test only), “Dell PowerEdge R260 running Red Hat Enterprise Linux 9 as Docker hosts (Docker test only), and a WANulator host.

A Bridgeworks white paper on the results of the test explains: “Whether your use case is aggregation and analysis, training AI models or backup and recovery, data is only useful if you can access it. Many companies use IBM Storage Scale to access their data remotely, often across long distances.”

“Distance introduces latency into the connection and increases the chance of packet loss. Both latency and packet loss significantly slow transfer rates, and it can make day-to-day data access a painful process.”

They claim that, even in the most challenging scenario they tested, with 300ms of latency and 0.1% packet loss, PORTrockIT was able to complete backup jobs 50 times faster than a traditional network architecture. This is achieved by using AI, machine learning and data parallelisation.

The company’s IBM Storage Scale white paper explains: “Parallelisation effectively eliminates the effects of latency by ensuring that the physical connection is constantly transferring new data packets from the sender to the recipient: there is no longer any idle time, and the network’s bandwidth can be fully utilised.”

Performance tests

To demonstrate the kind of results that PORTrockIT can deliver for IBM Storage Scale customers, Bridgeworks conducted a set of performance tests at a UK-based testing facility. The concept was initially explored on open-source virtualisation platform Proxmox, running PORTrockIT in Docker containers for rapid deployment. When the results showed promise, further tests were performed, using VMware ESXi for virtualisation and hardware PORTrockITs to achieve maximum performance.

The paper says the test infrastructures replicated a real-world WAN transfer, using a WAN emulator (WANulator) to simulate different levels of latency and packet loss between the source and the target systems. The tests transferred data between two clusters, each with a gateway and a pair of NSD servers. Each NSD server was equipped with two high-performance solid-state drives (SSDs) to ensure that the results were not bottlenecked by the write speed of the storage media.

The only limitations on the transfer speed were the 10Gb/s Ethernet link and the simulated conditions of the WAN. Each cluster had a local IBM Storage Scale filesystem and was configured to remotely mount the other cluster’s filesystem, providing read and write access. Transfers were performed using the IO500 storage benchmark suite, specifically the IOR-easy-write test. The IO500 suite is an industry standard tool for assessing the performance of filesystems, so it was ideal for this comparison.

The first set of tests involved unaccelerated infrastructure – directly connected to the WANulator. Then the same tests were conducted two hardware Bridgeworks PORTrockIT appliances, connected to either side of the WANulator, between the cluster and the WAN. For a final comparison, tests were also performed with two types of virtual PORTrockIT.

The results of the POC  

The results show that performance on the unaccelerated architecture degraded as soon as even a small amount of latency was introduced. Just 50ms of latency reduced performance from 1,000MiB/s to 319MiB/s, and when latency rose above 100ms, transfer rates dropped below 100MiB/s.

This effect was even more extreme when a small amount of packet loss was introduced. With 0.1% packet loss, it takes just 20ms of latency to bring the transfer speed down from 1,000MiB/s to 34MiB/s. However, when accelerated with WAN Acceleration, a different picture emerges.

With 0% packet loss, PORTrockIT maintains 938MiB/s even at 100ms of latency, while with 0.1% packet loss it reaches 450MiB/s when the unaccelerated transfer is struggling to reach 5MiB/s. Meanwhile, the 0.1% packet loss and 20ms of latency that previously crippled the transfer now barely has an effect, as PORTrockIT increases the speed by 25x.

The POC demonstrates that, “for typical packet losses of <1%, latency is by far the greatest determiner of bandwidth loss. However, even with no latency, packet loss can significantly harm the effectiveness of a connection. IBM Storage Scale has the ability to cope with up to 0.1% packet loss in a latency-free environment, but performance decreases rapidly thereafter. But by introducing PORTrockIT, we can significantly mitigate the impact of higher levels of packet loss.”

It also proves that organisations can benefit from AI and, in particular from AI-enabled mitigation, with WAN Acceleration. Its performance gains can boost productivity, the ability to back and restore data, while improving cloud and collaboration performance and demonstrating that AI can be a force for good - helping organisations to support service continuity and cut dataloss exposure.