Advanced scalable data infrastructure for organizations of all sizes

A number of GraphStream's clients are working to make incremental improvements to their existing datacenter network infrastructure. In order to minimize risk, these improvements are typically introduced and tested at a small scale, in one or more pods of servers and switches, and then gradually deployed more widely.

Opportunities for upgrading existing datacenter networks include:

  • Migration away from access-aggregation-core architectures, and toward Clos-type leaf-spine topologies with lower oversubscription ratios. This yields many benefits, including radically better scalability and lower cost per unit of capacity; greater flexibility in workload placement; and improved fault tolerance.
  • Improved automation and coordination of the control plane across a scalable fabric of physical and virtual network devices, via the introduction of enhanced abstraction layers, control protocols, APIs, and logically centralized controller software.
  • Migration of certain network functions away from non-scalable dedicated hardware+software appliances, and toward scalable, resilient services distributed across a cluster of virtual machines running on a combination of COTS general-purpose servers and network switches.

GraphStream is working together with clients to pilot and then deploy these types of incremental network upgrades, using software from multiple suppliers.

Many physical network switch products are available for purchase only as integrated hardware+software appliances. These appliances typically include general-purpose processors; specialized programmable hardware to accelerate packet forwarding; arrays of physical network ports; a Linux-based operating system; and an application-level software stack that implements a variety of network services.

Some physical network switch products are available for purchase in an alternative "bare-metal" format, with the same hardware but without pre-installed OS or application software. The purchaser then installs OS and application software packages selected from the offerings of multiple suppliers. This bare-metal model for network switches mirrors the separation of hardware and software choices that has long been available for general-purpose servers with x86-architecture processors. For many use cases, the bare-metal format is becoming a very compelling alternative.

For client engagements that include the deployment of hardware capacity for networking, GraphStream uses COTS interconnect hardware products as building blocks to deliver this capacity. We can deliver a variety of network hardware platforms in response to specific client requirements, but the majority of the network hardware that we are currently deploying is from the suppliers listed below.

The following suppliers provide key building blocks for the types of network upgrades described above.

Dell was the first Tier-1 data infrastructure supplier to introduce a line of Ethernet switch hardware products in the bare-metal format. Dell offers these products under its Open Networking brand, with support for network software stacks from multiple suppliers including Big Switch Networks and Cumulus Networks.

Mellanox Technologies offers an extensive line of software and hardware products for highly scalable InfiniBand network fabrics. The software products include tools for acceleration of scientific and engineering application workloads; and a full stack of enabling software for operating, monitoring, and troubleshooting InfiniBand fabrics. The hardware products include host bus adapters, and switch and cabling systems.

Use cases where InfiniBand networking is preferred over Ethernet networking include:

  • Installations that incorporate system software, middleware, and/or application software that is written to use InfiniBand APIs.
  • Applications that require network latency that is lower than what currently available Ethernet implementations can deliver.
  • Applications for which currently available InfiniBand implementations can meet latency and throughput requirements more cost-effectively than currently available Ethernet implementations can.