Test of the New Infortrend CS Scale-Out NAS Cluster (Part 1)

Scale-Out vs. Scale-Up

Compared to traditional scale-up architecture, scale-out storage systems allow users to add controllers at the same time of adding storage inventory. Thereby system processing power can grow linearly whilst the capacity expands.

In scale-out systems, instead of adding JBODs behind existing controllers, we add nodes to a cluster. A storage node has both storage function and computing function, which eliminates the bottleneck of total processing power in scale-up systems.

Digistor provides solutions with a number of the most well-known scale-out storage solutions, including Dell EMC PowerScale & Isilon, EditShare EFS and Qumulo. These systems meet our customer's production requirements, and they allow customer to:

• Scale elastically with low marginal cost and minimum downtime.
• Independently scale performance and capacity.
• Have consistent node performance.
• Have almost unlimited storage lake capacity within the same namespace
• Have more flexible control of protection levels.
• Have better operational experience.

The Infortrend CS System

Infortrend just entered scale-out market with their newly designed CS series of storage nodes. We recently received a 3-node CS 4024G cluster from Infortrend to test with M&E workflows. I would like to share my testing progress and results here in this series of blogs.

There are a few interesting design in the Infortrend CS system, which can rarely be seen in other competor systems:

1. Mixed scaling models
   The CS system supports both scale-up and scale-out in their scale-out architecture, users can deploy the system with a mixture of both scale methods in a single cluster.

1. Data protection with both RAID and erasure code
   When Isilon first invented scale-out NAS, their method of purely using erasure code (128KB FEC blocks) is adopted by almost all scale-out storage systems. This approach demands a massive amount of computing power during calculating and recalculating jobs, because the FEC algorithm can only execute with general CPUs. Infortrend chooses the way of mixing intra-node RAID and inter-node erasure code to accelerate the data protection process. RAID controllers process parity bits and data distribution with hardware-based ASIC and FPGA/CPLD, while CPU resources handles the erasure code for node redundancy. This application of RAID is also the reason why CS supports scale-up in its scale-out architecture.

1. Distributed mode for throughput-hungry environment
   Infortrend's 'distributed mode' writes only one copy of a file to one node only (node is selected dynanically based on hashing algorithm). Files are protected by RAID only without erasure code, thus the system can deliver better throughput. This mode is suitable for environments demand heavy I/O over extreme protection.

Infortrend CS 4024G 3-Node Cluster Setup

The CS 4024G nodes support 10/25/40GbE networks in both front and back ends, and is capable of handling multiple HD and 4K streams in large-scale editorial and postproduction workflows.

Our workshop test includes the below hardware components:

• 3 x Infortrend CS 4024G nodes
• 72 x 12TB Seagate SAS drives
• Mellanox SN2100 100GbE fibre switch
• Dell 5548 1GbE switch

Back-end and Management Network

In our setup, each node uses two 25GbE connections to the back-end switch (Mellanox SN2100) and the management interfaces are connected to a seperate management switch (Dell 5548).

Front-end Network

The core switch of the front-end network is the same Mellanox SN2100, I use different VLANs to separate back-end and front-end traffic. The access switch where client machines connect to is a Dell EMC S4128F-ON, with an 100GbE uplink to the core Mellanox.

Each node has 6 front-end connections: four 10GBase-F and two 25GBase-F. I setup LACP for each port group to simplify the management and round-robin access load-balancing.

Storage Pool

After the initialisation of each node, and setup a protection level, the system presents us a storage pool with 684TB usable capacity. For the first test, the protection level is setup as:

• RAID5 on each node (optimal performance)
• Erasure code 2+1 for the cluster

While I am still running a set of benchmarks on the system, just to quick show you how the cluster looks in the rack:

In my next blog, I will share some of my test results, meanwhile, feel free to leave a message and let me know what you are expecting to see in the benchmark. I am happy to include a few extra tests based on the comments.