Datrium has added an “Insane Mode” feature that the startup claims can double host storage performance on the fly.
Datrium took the Insane Mode tag from Tesla’s rapid acceleration technique for its Model S cars. The concept is similar – the goal of Datrium’s Insane Mode is to let customers instantly speed the storage systems performance by increasing available CPU resources.
Here’s how Datrium DVX works:
The system has two parts. The Diesl (Distributed Execution Shared Logs) file system runs on the host and NetShelf storage array. Diesl allows customers to manage storage like virtual machines through VMware vSphere, and handles features such as compression, deduplication and instant zero-copy clones. It uses up to 8 TB of NVMe flash in the server to boost performance. Customers can use any commercial flash in the host.
NetShelf is the second part, and serves as durable storage. NetShelf has dual controllers – each with NVRAM – for active/passive failover and hard disk drives for capacity. It exports no blocks or files and has no LUNs or volumes to manage. Diesl provides host-based data services. Writes are mirrored before control is returned to a virtual machine. A failed host does not impact fault tolerance or cause data loss. A NetShelf includes 29 TB of data after RAID 6 overhead.
Insane Mode an alternative to QoS
Insane Mode is the third method for improving DVX performance. The others are to add flash to the host or to vMotion a workload to another host with more available flash.
Insane Mode, now part of Diesl Hyperdriver Software, reserves 40% of host cores for DVX instead of the default 20%. Datrium recommends using Insane Mode only when a host averages less than 40% CPU utilization for VMs because ESX utilization should be kept below 80%. Customers can stay in Insane Mode indefinitely or use it for specific tasks like running batches before returning to 20% utilization. Customers can go from Fast Mode (the default) to Insane Mode with one mouse click from the DVX dashboard.
Datrium marketing VP Craig Nunes said doubling the CPU resources can improve workload performance from 1.5x to 3x “literally instantly.”
He said the feature is an alternative to quality of service on traditional arrays. “A storage array has a set amount of performance in its controller. When you hit the boundaries, it’s upgrade time,” he said. “Quality of service on the array is about how you provision precious resources in a storage array. The approach we take is, service levels are more about giving more resources for performance instead of limiting precious resources.”
Datrium CTO and founder Hugo Patterson said the extra CPU is available because customers rarely go above 50% utilization. “If you’re not doing I/O, then CPUs are available for other things,” he said. “This basically reserves enough so when you need to use it, it’s there.”
Patterson said he expects Insane Mode to be used for mainly for test/dev, virtual desktop infrastructure and server consolidation.
Datrium bucks all-flash, hyper-converged
Patterson sees Datrium as a superior to two hot storage trends – all-flash arrays and hyper-converged systems.
“Flash belongs on the host, and flash has been on the host for a decade,” Patterson said. “But when it is on the host, it’s just a storage device. It’s an island of flash, not an enterprise class storage system. One SSD is not an enterprise class storage system. So how do you get flash on the host and yet make it an integral part of an enterprise storage system? That’s what the DVX is really all about. We’ve moved a lot of the storage compute – just about all of it – into the host. These days it’s a bigger deal to do that with the storage compute getting bigger and bigger.”
Patterson also argues that a DVX costs less and uses both the server and storage better than a typical hyper-converged system.
“Each server is independent [with DVX] and gives an easier way to manage storage performance,” he said. “Communication is from servers to NetShelf and servers don’t talk among themselves, except in the case of vMotion. We don’t have one host accessing data that’s on another server. When you have a cluster of [hyper-converged] servers, they’re all co-dependent and they can heavily influence the performance of each other. When performance isn’t what you hoped for, it’s hard to do anything about it. With DVX, each server has its own local cache and uses its own CPU.”