Rakuten targets big edge deployments with Symworld Cloud

Abstracting hardware allows for 100% utilisation

“Our strength in relation to the big Kubernetes players – Red Hat OpenShift and VMware Tanzu – is that our system was conceived of early on for very large deployments,” said Partha Seetala, president of Rakuten Symphony, to Computer Weekly’s French sister publication, LeMagIT, during a recent IT Press Tour event.

“Our competitor’s approach is to manage application instances in a uniform hardware cluster with one type of CPU, GPU, storage and network card,” said Seetala. “Ours is exposed in the same way from different hardware so that applications get the maximum performance as if written expressly for them.

“Take for example these case in telecoms, when an OpenRAN deployment wants to validate a communication in less than 40μs. It is simply impossible to do if your software code doesn’t take into account the type of network or FPGA accelerator in your equipment. But you have 800 hardware variants in your antenna park. What do you do? Write and deploy 800 versions of your code? That’s what you have to do with our competitors. We, however, expose our underlying hardware in a universal manner.”

Rakuten Symphony’s software offer currently comes in three products. Symworld CNP (Cloud Native Platform) is the Kubernetes system bought from Robin.io. Symworld Orchestrator is the console for remote admin of edge servers, something like VMware vCenter but capable of managing 100,000 physical machines and more than 10,000 Kubernetes clusters.

Finally, Symworld CNS (Cloud Native Storage) corresponds to the storage layer in Robin.io CNP, which can be installed on top of any other Kubernetes system, with OpenShift particularly in mind. According to the company, Ceph storage, which Red Hat supplies for use with Kubernetes, is not completely suited to edge deployments.

CNS: To go beyond CSI

“When it comes to storage, we enhance Kubernetes with high-level functions similar to those you’d find in storage arrays for virtual machines,” said Seetala. “At the edge you typically deploy applications that use MongoDB. The first thing is to enhance Kubernetes so it has persistent storage. That’s what the array makers do with CSI. But you have to go further.”

Seetala praised the characteristics of its storage layer, which manages bandwidth in multiple containers on-the-fly, tiering between drives according to workload, encryption and compression.

“Above all, our system protects data, makes local snapshots, backups elsewhere, and offers a complete system to restore activity in case of incidents, including via content synchronisation between servers or clusters.”

All of which takes into account the performance characteristics, bandwidth, IOPS and disk capacity on each server.

Regarding backup and snapshots, Seetala explained it isn’t enough to put an emergency copy of MongoDB on disk. To restore a functional instance you have to capture transactions as they happen, such as application configuration – especially pipelines between the various services – and metadata.

CNP: The flagship

The storage layer works with a CSI driver, supplied by CNP for internal disks or by the disk array maker. And CNP is not autonomous, being installed on a Linux tier. Rakuten Symphony uses CentOS and RHEL from Red Hat.

A key component of CNP is its Sherlock module, which diagnoses host hardware – Rakuten Symphony works with key chips and server suppliers – and produces a pool of resources. These comprise the performance of the processor and other accelerator chips, the network and the storage.

This performance pool is exposed to Kubernetes via the Advanced Scheduler module, which allows the orchestrator to deliver resources according to application needs. To make the most of the pool of resources, applications can use a graphical tool to help specify their needs, such as network bandwidth, CPU or GPU power.

Sherlock and Advanced Scheduler are packaged as part of the CNS product but are limited to discovery and sharing of storage resources.

Advanced Scheduler can distribute resources to any virtual machine. “We know well that containers don’t 100% replace virtual machines. So, if your edge server runs VMs from a Linux KVM host, it is treated by CNP as it would containers,” said Seetala.

The network layer is provided by Calico, the open source system used in OpenShift. It isn’t a mesh network that allows containerised applications to send messages, but a TCP/IP network that provides functionality as firewall and virtual network card per container.

A graphical console for remote supervision and monitoring

Finally, Symworld Orchestrator brings deployment, updates and remote monitoring functionality similar to that found in management consoles in VMware or Nutanix.

It manages its fleet of machines using functional or geographic tree-like visualisation, with manual control or automation via customisable rules possible. It allows customers to run deployment scenarios that consist of, for example, deploying master servers locally and dependent servers in further locations.

It’s also Symworld Orchestrator that manages all operations connected to storage and data protection. CNS is delivered with a reduced version of the console to enable definition of rules for backup and recovery of activities.