VAST Data – No More Tiers Means No More Tears?

Disclaimer: I recently attended Storage Field Day 18.  My flights, accommodation and other expenses were paid for by Tech Field Day. There is no requirement for me to blog about any of the content presented and I am not compensated in any way for my time at the event.  Some materials presented were discussed under NDA and don’t form part of my blog posts, but could influence future discussions.

VAST Data recently presented at Storage Field Day 18. You can see videos of their presentation here, and download my rough notes from here.

 

VAST Enough?

VAST Data have a solution that basically offers massive scale with Tier 1 performance, without the cost traditionally associated with Tier 1 storage.

Foundational Pieces

Some of the key pieces of the solution are technologies that weren’t commonly available until recently, including:

  • NVMe-oF – DC-scale storage protocol that enables remote NVMe devices to be accessed with direct attached performance.
  • QLC Flash – A new Flash architecture that costs less than enterprise Flash while delivering enterprise levels of performance.
  • Storage Class Memory – Persistent, NVMe memory that can be used to reliably buffer perfect writes to QLC and create large, global metadata structures to enable added efficiency.

If you read their blog post, you’ll notice that there are some interesting ideas behind the VAST Data solution, including the ideas that:

  • Flash is the only media that can be used to bring the cost of storage under what people pay today for HDD-based systems.
  • NFS and S3 can be used for applications that up until now required a level of performance that could only come from block storage.
  • Low-endurance QLC flash can be used for even the most transactional of workloads.
  • Storage computing can be disaggregated from storage media to enable greater simplicity than shared-nothing and hyper-converged architectures.
  • Data protection codes can reduce overhead to only 2% while enabling levels of resiliency 10 orders of magnitude more than classic RAID.
  • Compressed files provide evidence that data can be reduced further when viewed on a global scale.
  • Parallel storage architectures can be built without any amount of code parallelism.
  • Customers can build shared storage architectures that can compose and assign dedicated performance and security isolation to tenants on the fly.
  • One well-engineered, scalable storage system can be ‘universal’ and can enable a diverse array of workloads and requirements.

Architecture

[image courtesy of VAST Data]

  • VAST Servers – A cluster can be built with 2- 10,000 stateless servers. Servers can be collocated with applications as containers and made to auto-scale with application demand.
  • NVMe Fabric – A scalable, shared-everything cluster can be built by connecting every server and device in the cluster over commodity data center networks (Ethernet or InfiniBand).
  • NVMe Enclosures – Highly-Available NVMe Enclosures manage over one usable PB per RU. Enclosures can be scaled independent of Servers and clusters can be built to manage exabytes.

Rapid Rebuild Encoding

VAST codes accelerate rebuild speed by using a new type of algorithm that gets faster with more redundancy data. Everything is fail-in-place.

  • 150+4: 3x faster than HDD erasure rebuilds, 2.7% overhead
  • 500+10: 2x faster than HDD erasure rebuilds, 2% overhead Additional redundancy enables MTBF of over 100,000 years at scale.

Read more about that here.

Global Data Reduction

  • Data is fingerprinted in large blocks after the write is persisted in SCM
  • Fingerprints are compared to measure relative distance, similar chunks are clustered
  • Clustered data is compressed together; byte-level deltas are extracted & stored

Read more about that here.

Deployment Options

  • Full Appliance – VAST-provided turn-key appliance
  • Software-Defined – enclosures and container software
  • Software-only – run VAST SW on certified QLC hardware

 

Specifications

The storage is the VAST DF-5615 Active / Active NVMe Enclosure.

[image courtesy of VAST Data]

 

I/O Modules 2 x Active/Active IO Modules
I/O Connectivity 4 x 100Gb Ethernet or 4 x 100Gb InfiniBand
Management (optional) 4 x 1GbE
NVMe Flash Storage 44 x 15.36TB QLC Flash
NVMe Persistent Memory 12 x 1.5TB U.2 Devices
Dimensions (without cable mgmt.) 2U Rackmount

H: 3.2”, W: 17.6”, D: 37.4”

Weight 85 lbs.
Power Supplies 4 x 1500W
Power Consumption 1200W Avg / 1450W Max
Maximum Scale Up to 1,000 Enclosures

 

Compute is housed in the VAST Quad Server Chassis.

[image courtesy of VAST Data]

 

Servers 4 x Stateless VAST Servers
I/O Connectivity 8 x 50 Gb Ethernet 4 x 100 Gb InfiniBand
Management (optional) 4 x 1GbE
Physical CPU Cores 80 x 2.4 GHz
Memory 32 x 32GB 2400 MHz RDIMM
Dimensions 2U Rackmount

H: 3.42”, W: 17.24”, D: 28.86”

Weight 78 lbs.
Power Supplies 2 x 1600W
Power Consumption 750W Avg / 900W Max
Maximum Scale Up to 10,000 VAST Servers

 

Thoughts And Other Reading

One of my favourite things about the VAST Data story is the fact that they’re all in on a greenfield approach to storage architecture. Their ace in the hole is that they’re leveraging Persistent Memory, QLC and NVMe-oF to make it all work. Coupled with the disaggregated shared everything architecture, this seems to me like a fresh approach to storage. There are also some flexible options available for deployment. I haven’t seen what the commercials look like for this solution, so I can’t put my hand on my heart and tell you that this will be cheaper than a mechanical drive based solution. That said, the folks working at VAST have some good experience with doing smart things with Flash, and if anyone can make this work, they can. I look forward to reading more about VAST Data, particularly when they get some more customers that can publicly talk about what they’re doing. It also helps that my friend Howard has joined the company. In my opinion that says a lot about what they have to offer.

VAST Data have published a reasonably comprehensive overview of their soilution that can be found here. There’s also a good overview of VAST Data by Chris Mellor that you can read here. You can also read more from Chris here, and here. Glenn K. Lockwood provides one of the best overviews on VAST Data you can read here.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.