Fusion-io has debuted the ION data accelerator, a Linux-based software stack that turns the ioMemory within a server into a shared storage target.
Fusion-io executive David Flynn noted that VMware’s recent acquisition of Nicira had brought software-defined networking into the spotlight. ION, Flynn claimed, inaugurates vendor-neutral, software-defined flash storage: “The biggest compliant that people have had is that with flash attached as a memory within a server, you don’t get the benefits of sharing between servers.”
The technology “turns the server into something that looks like the highest-performing SAN on the planet,” he added. “We’re talking about a 1U server that can do over a million I/Os a second and over 6 Gbytes per second of bandwidth over standard networking, from off-the-shelf software.”
However, ION isn’t the first attempt in this space: both Violin Memory and Texas Memory have adopted similar approaches (Fusion-io claims that both rivals’ solutions are proprietary). Texas Memory recently announced the development of bootable PCI Express-based flash storage, an unrelated development but one that the company claims will alleviate the need for traditional hard drives within arrays; just SSDs may be used, the company claims.
Customers have been asking for Fusion-io memory that’s accessible over a network, but Flynn said the company held off until it could do so with satisfactory performance. In order to help make that happen, companyhired several key Linux authors, including Jens Axboe, the current Linux kernel maintainer of the block layer; Chris Mason, the author of Btrfs and Fusion-io’s director of kernel engineering; and others.
The ION can deliver 600,000 8K IOPS, with 5 GB/s of bandwidth. Latencies are just 73 microseconds for reading and 56 microseconds for writing. Up to 20 Tbytes may be addressed, over either Fibre Channel or Infiniband.
Fusion-io, which debuted the software Aug. 1 in conjunction with an event in San Francisco, will charge $3,900 for the software, Flynn said. That price is for the server license, regardless of the ioMemory used with it. That requires the customer to integrate the software with an existing server that uses ioMemory; Cisco, Dell, HP, IBM all provide those machines. In this case, the customer will need to solicit server support from the server vendor, ioDrive support from the supplier, and ION support from Fusion-io.
Customers can also purchase a fully integrated solution using hardware from either HP or Supermicro, and receive support just from Fusion-io itself.
“The interesting part about this is that by disaggregating, by decoupling our solution from the actual server box, it not only allows choice of vendor and the diversity of form factor,” Flynn said. “So you have this immense diversity of form factors so that it can fit into your homogenous array. If you like blades, it’s just another blade. If you like pizza boxes, it’s just another pizza box – it’s the same form factor.”
Image: Jakub Pavlinec/Shutterstock.com
Resiliency is always an issue with PCIe devices. You can't have one PCIe device accessed by two SAN controller heads (unlike SAS or FC interfaced SSD devices).
The only use I see for Fusion-IO is niche-market type applications where extreme speed is required, but 100% uptime is not critical, or as caching disks (although you have to be careful using these as write cache as this could lead to data loss if the server goes down in a shared primary storage environment). This is unfortunate as I really wanted to see a cool mainstream use for these as primary storage!
If you need mission critical storage for a reasonable price you could, for example, run a Nexenta storage cluster with huge amounts of RAM (which ZFS will use as ARC read cache) and multiple STEC ZeusRAM devices for write cache (with SAS interfaces). Then backend the whole thing with standard SAS SSD drives for mass storage.
This would give you a very fast clustered storage solution.
I think Fusion-IO needs to come out with some kind of PCIe redundant solution... Maybe a 1U box populated with Fusion-IO drives (should be all ASIC driven, another OS here would be a single point of failure), and two Light Peak-style interfaces (which is basically PCIe on a wire/fiber) to two separate storage controllers. The storage controllers would then run some kind of clustered file system (active/active) or ZFS-style filesystem (active/passive) with cluster resiliency between them. On the front-end the controllers could run FC or iSCSI to the hosts (a partnership with Nexenta could work here, getting their "special" Light Peak drivers in to the Nexenta kernel).
- spam
- offensive
- disagree
- off topic
LikeAnd how come it's not proprietary if you have to se ION and IO-drive to make it work? This smells like a lash up because PCI solutions have limited applicability and use cases and FIO have a big stock price to support.
- spam
- offensive
- disagree
- off topic
LikeJust wanted to understand ,how will a failure of a server access the share storage ensuring no data loss ???
- spam
- offensive
- disagree
- off topic
Like