Pavilion Data has fulfilled a road map promise to add S3-protocol object storage to the RF100 multi-controller all-flash array
At launch last month the array – which Pavilion dubs a “Hyperparallel Flash Array” (HFA) – supported 20 controllers in the 4U enclosure. The controllers talk to NVMe SSDs across an internal PCIe network and the array delivers 120GB/sec read, and 90GB/sec write bandwidth, with 20 million IOPS at 40µsec latency. It supports access via NVMe-oF/TCP, NVMe-oF/ RoCE, NFS, iSCSI, Ethernet and InfiniBand.
Gurpreet Singh, Pavilion CEO, said in prepared statement yesterday: “With the addition of native fast object support to the Pavilion HFA, customers can, for the first time, enjoy both industry leading high performance and ultra-low latency from one platform in a small footprint.”
The S3 storage adds to the existing block and file access and enables tiering to the public cloud for longer term storage.
Performance
Pavilion said HFA S3 storage delivers 80 GB/sec read and 35 GB/sec write throughput. This is much faster than most object stores, certainly disk-based and dual-controller ones, but MinIO and OpenIO have both demonstrated greater speed.
OpenIO achieved 1.372 Tbit/s throughput (171.5GB/sec), using an object storage grid running on 350 commodity servers. MinIO exceeded that, delivering 183.2 GB/sec (1.46 Tbit/s) on reads and 171.3 GB/sec (1.37 Tbit/s) on writes. The company thinks it can achieve 2.92Tbit/s read bandwidth – 366GB/sec.
HW-boosted MinIO
Pavilion is using a derivative of MinIO object software technology, “optimized to leverage our IP of our overall Hyperparallel Storage Solution,” director of Product Marketing Keith Parker told us.
The Minio system above used a 32-node set up. Parker told us: “Pavilion would need to offer only 3 Pavilion HFA arrays consuming 12U of rack space, which would give Pavilion 240Gb/sec read.”
He says Pavilion is also faster than MinIO in AWS when compared on a per-rack unit basis: “A testament to our integration/implementation of MinIO is the performance we get Per RU relative to MinIO’s published numbers on AWS.
“If you normalise this to single Rack Unit (assuming the AWS servers they are using are 2U) Pavilion is; 7x the read performance, 3.3x the write performance [and] 8x the PB capacity. This is all using “the same software”, so clearly our implementation and integration to our hardware makes a material difference.”
In other words MinIO SW is fast but Pavilion HW makes it go faster still.
Phased namespace extension
The initial Pavilion S3 implementation is limited to one S3 namespace per controller. Phase 2 will see a single S3 namespace across multiple controllers, and phase 3 extends that to a single namespace across multiple HFA systems.
The S3 data can be accessed via S3, only not by any file interface. Other object storage suppliers, such as Cloudian and Scality, have added file access to their object storage.
