"iSCSI killer" Native in Linux

jar writes "First came Fibre Channel, then iSCSI. Now, for the increasingly popular idea of using a network to connect storage to servers, there's a third option called ATA over Ethernet (AoE). Upstart Linux developer and kernel contributor Coraid could use AoE shake up networked storage with a significantly less expensive way to do storage -- under $1 per Gigabyte. Linux Journal also has a full description of how AoE works." Note that the LJ article is from last year; the news story is more recent.

AOE?

I didn't know Age of Empires can do network storage! WTG Microsoft!

Will it catch on?

From TFA:
Some significant caveats mean that not everyone is so keen on the technology. For a start, it's a specification from Coraid, not an industry standard. Its networking abilities are limited. And its detractors include storage heavyweights such as Hewlett-Packard and Network Appliance.

So will this ever develop into a real standard or will it remain the sole domain of one company? I do not know if I want to invest time and money into it if the latter is true. From a comp sci point of view this is a great approach to networked storage. It uses what people already have to make storage reletively cheap. I am going to wait to see where this technology goes. Maybe it will blossom and become a serious contender.

Re:Will it catch on?

For a start, it's a specification from Coraid, not an industry standard.

I don't know that this is true, because the LinuxJournal article directly contradicts it. (Unless I'm misreading it.) Here's what the LJ says:

ATA over Ethernet is a network protocol registered with the IEEE as Ethernet protocol 0x88a2.

So, it looks like the protocol has been officially registered and was granted approval by the IEEE--so that makes it an industry standard. It may not be adopted yet, but it's certainly not something like 802.11 pre-n or anything; there's an official and approved protocol.

Re:Will it catch on?

So, it looks like the protocol has been officially registered and was granted approval by the IEEE--so that makes it an industry standard. It may not be adopted yet, but it's certainly not something like 802.11 pre-n or anything; there's an official and approved protocol.

Anyone can register a protocol number with IEEE by paying a $1000 fee. It doesn't mean it's a protocol endorsed by IEEE in any shape, way or form.

Re:Will it catch on?

The non-routable is a killer. Protocol-level bridging, no off-site redundancy, strict dependancies on port location. No thanks, it's a toy protocol that may get some use in the home NAS market, but it was hell to implement a reliable setup in our lab un

Cheaper?

I guess I don't really see how it's cheaper that iSCSI? Sure, there's less overhead from the lack of TCP/IP, so you may not need as massive a network to drive it equally. But I've been under the understanding that iSCSI doesn't require SCSI drives, so you could build an iSCSI target out of the same machine/drives as an AoE host, correct? For some applications, I think the lack of TCP/IP might be a benefit - less opportunity to hack. (Then again, I'd expect anybody deploying something like this or iSCSI would drop the few extra $$$ to build a parallel network that transports just storage.)

Re:Cheaper?

The main advantage of AoE is that it's simple enough that you could build it in hardwired silicon if you wanted to, or use small microcontrollers way smaller than what you'd need to run a fullblown TCP stack (this is what Coraid does, I believe.)

The main d

Re:Cheaper?

I hacked together an iSCSI setup from some old hardware.

2 P II 400MHz systems running FC4
One system had software raid 0 on 2 IDE drives.
The target has a spare 10GB IDE drive.

Added 2 10/100T cards with a crossover cable.

Did a quick dd if=/dev/zero count=som

Re:Cheaper?

The really silly thing about this is that they claim it's "lower overhead" than TCP/IP because people are having to buy "expensive TCP offloading engines" for iSCSI, when a few seconds of research provided, namely on Wikipedia (http://en.wikipedia.org/wiki

Yes!

I like the look of this technology. The great thing it has going for it is that most of the non-hard-disk infrastructure (switches and cabling) leverages the tremendous investment in ethernet. That is great.

The thing that needs work, in my view, is that the bit that links the disks and the rest isn't cheap enough. In fact what would be awesome here is if, say, Seagate provided disks with native ATAoE connectors built-in. They might have to buy Coraid for that to happen.

In case anyone thinks I'm out of my mind here, don't forget that disks can already be had with ATA interface, SCSI interface, FCAL interface, SATA, SAS - that's five and there are probably more. Yes they might be a bit more expensive, but if they come in under the combined price of "regular ATA disk" + Coraid ATAoE disk adapter then you'd come out ahead. Someone like Seagate would, I think, have the industry-wide clout and respect to succeed in making this an open standard. Something that will be a challenge for Coraid for a long time (I have nothing against them, btw, they are friendly and their mailing list didn't spam me when I signed up).

When I was on the OpenSolaris pilot project I tried to get people interested in using this with Solaris. I think it might be great for ZFS, for example. At that point the real storage wizards were more interested in iSCSI, but I respectfully disagree, OpenSolaris + ZFS + cheap storage = awesome file server. Emphasis on the cheap. As Sun people will admit, their previous attempts at RAID were more like RAVED (Redundant Array of Very Expensive Disk). Coraid does have a Solaris driver, so this is definitely feasible.

Re:Yes!

I like the look of this technology.

    It's the eyeliner. It doesn't look half as good in the morning.

iSCSI killer?

In the context of using this in low-cost environments with Linux I can hardly see how this could kill iSCSI. Last week I implemented an iSCSI setup for about 500 euros (target serves out 500GB disk space for non-critical backup) using standard components, FC5, iSCSI Enterprise Target [sourceforge.net] and Microsoft iSCSI Initiator.

Works great and is a lot (>10x) faster than the about similarly priced NAS device that was used for the same task before.

Not an iSCSI killer, here are the reasons why not

1) Complexity for RAID and volume management is not centralized and is pushed to individual hosts. One of the main benefits of SAN technology is that you can just carve out storage from a single interface and assign it to a server and the server simply sees it as a block device. With AoE each drive is addressed separately by the server, which means it is up to the server (and server admin) to figure out how to handle distributing over multiple drives, handle drive failures, and expanding volumes. This is huge.
2) It is not a standard and is only really supported by one vendor. This may change in the future but it is significant right now. It is registered with the IEEE but that hardly makes it a peer-reviewed standard with input/improvements from many experts.
3) No boot from SAN. Until someone makes some sort of mini bootstrap system on a CD or a hardware card implementation of AoE that can be addressed as a block device admins will be unable to host the root filesystem and/or C: drive on an AoE SAN
4) No multipath (that I can see). Perhaps I misunderstand this, but it seems like there is no way to do multipath IO with this system. That is, all the drives are single-connected to a network. If that network switch goes down, all drives on that network are inaccessible.
So AoE looks like a neat technology for pushing drives out of the box and potentially sharing them among hosts, but there is no intelligence there. It is just dumb block addressable storage with no added availability or management, and therefore is far from being an iSCSI or FC killer.

Re:Not an iSCSI killer, here are the reasons why n

You can use Ethernet-based multipath IO, a lot of switches can be stacked to provide redundancy (and load-ballancing).

AoE is a COOL thing exactly because it's a 'dumb' technology. You can buy a switch, a bunch of disk drives and AoE adapters, a small Linux PC - and your storage system is ready. There is a lot of existing RAID manipulation and monitoring tools for Linux, so RAID configuration is not a problem.

You also can boot from SAN, it's not a problem. Just add required modules and configs to initrd and place it on a USB drive.

ATAoE is a crock, it's no better than iSCSI

So. Coraid has not, in a whole year, explained why iSCSI is somehow more expensive (disks + Linux kernel + network.. all the same) than their ATAoE implementation.

They'll give excuses about the cost of iSCSI hardware offload.. but you don't need that. ATAoE is all software anyway it's just a protocol over ethernet, rather than layered on top of TCP/IP.

What is wrong with using TCP/IP - which is already standard and reliable? Nothing. We know TCP/IP provides certain things for us.. resilience (through retransmits), and routing, are a good couple, and what about QoS?

ATAoE needs to be all the same network, close together, they're reimplemented the resilience, you can't use inbuilt common TCP checksum, segmentation and other offloads in major ethernet chipsets because they're a layer too low for it.

No point in it. Just trying to gain a niche. They could have implemented products around iSCSI, gotten the same performance with the same features, for the same price. Bunkum!

I can't tell if this is clever or stupid...

A wise man once told me there is a fine line between them.

ATA is a crappy protocol, even when local. It's only good for squeezing that last $0.03 out of the controller cost. Once you are using ethernet cables ($1) and links and PHYs on each end ($4 each), it makes a lot more sense to put some brains back in. Use SCSI. Heck, even ATAPI optical drives (the optical drive in your computer) uses ATAPI, which is SCSI in packetized ATA transfers.

Also, I'm a bit nervous about the packet CRC validation being done in the ethernet controller/layer itself. The problem is that if an ethernet switch between you and the storage device stores packs and forwards them (as all smart switches do), it may also chose to regenerate the CRC on the way. If it corrupts the packet internally and generates a new, valid CRC for the new, corrupt packet, you have undetected corruption. I'd be a bit nervous about that for my hard drive.

I do think using GigE is a smart way to attach hard drives to servers. I look at the back of an Apple XServe and see two GigE ports and a fibre channel card. Why can't one GigE port be used to attach to the network and one to the XServe RAID? Why do I need to get a multi hundred dollar card to attach to the XServe RAID when that GigE port is fast enough? It'd sure save a lot of cost, and hopefully reduce the price ot the end user.

Anyway. I'm pro GigE attachment, not sure I'm for this AoE.

I just deployed an AoE SAN

AoE rocks. It is very easy to set up, way simpler than iSCSI or fibrechannel or any other SAN technology I have used. And it enabled us to have many more options for high availability or clustered filesystems (which we are not yet using but I have been following the progress of GFS and Lustre, learning towards Lustre). We did not buy the Coraid stuff but instead used vblade on our own disk machines. A disk node in our cluster has 4 300G SATA disks which we RAID 5, 512M RAM, and the cheapest CPU Intel currently makes. We have dual core Opterons with 4G of RAM each with no internal disk. They PXE boot and then mount root straight off the AoE. Then we run Xen on the Opteron boxes. This is the killer setup. We can migrate xen domains avoiding downtime for hardware maintenance and if a machines dies we can instantly restart it on another machine because it all runs off the AoE SAN.

So far I am very pleased. Just make sure you get hardware that can do jumbo frames as this will increase your performance by 50%.

Re:Reliability

how is that relevant to the discussion of protocols?

reliability of SCSI versus ATA is largely imagined and the rest is intentional. drive manufacturers want you to believe their enterprise drives are more reliable and right now those drives are largely SC

Re:Reliability

People often forget there is a considerable difference in the reliability of ATA drives versus SCSI. If you are going to use some sort of ATA based SAN be prepared for disk failures much sooner than if they were SCSI.

This is not necessarily true. [storagereview.com] It all depends on how your network storage is being used. SCSI drives are built and firmware'd for the sole purpose of running a server, and they consistently beat any ATA drive (be it IDE or Serial) when it comes to server performance and reliability. ATA drives just aren't built to handle the sort of usage a server requires--note that this isn't a reflection of quality, but of purpose. But a file server (which is the only thing the SAN would be used for) requires much less robust firmware than a server housing MySQL, PHP, maybe a CRM suite, e-mail server, etc.--and so ATA drives shouldn't immediately be ruled as less reliable. The maturity of the technology plays a more important role than the interface.

Re:Reliability

the odds of 3 drives failing at once are astronomical.

No, they aren't. Just have an array running for a year or two and bring it down for maintenance, your chances of multiple drive failures are VERY good. Of course that happens even with SCSI drives, bu

Re:How does it lower costs?

Oops, only the linux journal article is down, the cnet article has answered my question: it isn't any cheaper than iSCSI + SATA solutions. $4,000 without any drives, compared to a starting price of $5,000 for a StoreVault (new from NetApp) with 1TB of stor

Re:How does it lower costs?

I think you are probably looking at the cost to buy Coraid's gear. You do not have to buy their stuff, although I am sure that they prefer that you do. I built my own AoE SAN using regular PC's. Way cheaper. I take the google approach: Use a larger amount of commodity hardware and design the system in an intelligent way to achieve the same performance and reliability at a better price/performance. Coraid hardware is basically just a Linux box with disks exporting AoE volumes. The nice thing about it is that you get their support. But AoE is so simple that you generally don't need support beyond perhaps the mailing list.

Re:Another "Killer"

AoE is a networked block device technology. NFS and Samba are network file system. One is about making block level access to a device available over the network, the other is about making file operations available.

In the case of AoE, a single remote block device can be shared between multiple systems. Each client could issue it's own write/reads. in combination with a distributed file system, each node could mount the same FS.

It's the same as NBD, iSCSI, Shared SCSI, and Fiber Channel.

Re:More for business?

Maybe cheapie little IDE hard disks are under $1/GB. If you want hot-swap, availability of half-decent RAID cards and disks which actually get to see some testing before they leave the factory, then you'll have to spend quite a bit more.

Re:More for business?

iSCSI is slightly differnet as rather than presenting a file system, it presents a hardware device. So you show it a 1TB device over the network (e.g /dev/sdb) then the client machine can partition that disk up as if it was local. Thats the advantage over