Slashdot Mirror
Wozniak Accepts Post At a Storage Systems Start-Up
Hugh Pickens writes "Apple co-founder Steve Wozniak is going back to work as chief scientist at Fusion-io, a start-up company that tweaks computers to let them tap vast amounts of storage at very quick rates. In the early days of Apple, Wozniak stood out as one of Silicon Valley's most creative engineers, demonstrating a knack for elegant computer designs that made efficient use of components and combined many features into a cohesive package and Wozniak will do similar work at Fusion-io, although this time with larger server computers and storage systems rather than PCs. 'I have a pretty quiet life, and I like to watch technology evolve,' says Wozniak. 'In this case, I like the people and the product, and said I would like some greater involvement.'"
It's always good to stay busy, and doubly so if you can actually do something that helps grow the existing technology.
And if he can make some cash from this gig, even better!
Go Woz!
Sounds like a pretty simple idea, I don't know why it hasn't been done before.
I've had the chance to play with some pretty phenomenal solid state drives (SSD) lately and, I have to say, that I can't believe that there isn't more industry buzz.
In a few months, an extra $100 will probably buy 120GB SSD, which will make a given PC perform like something completely different (you really need to go test drive an SSD PC if you have not yet indulged).
In a decade, I can see handhelds having so much storage and so much processing power, that we'll all just carry around our PC-on-a-phone and just use a standard interface to put that PC on any external monitor and keyboard. Hell, I can USB boot Ubuntu from my Blackberry, already.
More
Woz was always my hero. I was just a pimply faced kid when I first discovered Apple IIs (or more correctly, Apple II compatibles, since I was from a 3rd world country). Then I started reading about what he did, and his designs and so on. And when AAPL went public, he gave away his own shares to people who helped Apple get off the ground. Very very nice, very down to earth guy, from what I read about him. IIRC, he wanted to sell the Apple Is for $200 or so, and Jobs wanted $2000, and they settled on $666.66.
I was so disappointed when he left Apple and quit working on the Apple II series - that was such a great computer, and ahead of its times.
If its life depended on it, you might as well go to their site to learn what they do: http://www.fusionio.com/Products.aspx
"The module slides into certain slots inside servers. That gives the main computing chip quick access to data stored on the flash chips." I know that a general publication has to avoid jargon; but that sort of vague-but-vaguely-specific circumlocution looks like the server architecture equivalent of an awkward sex-ed class.
Good for Steve! The world needs more minds like his in the game.
Twenty years is a long time for a creative type to be mostly kicking back and enjoying the good things in life. Rust sets in quickly, dulling the drive that keeps one working nights and weekends and eating bad takeout food while crazy project deadlines loom. I wonder how much Woz has left.
For the company, it might not matter. Hiring Woz helps build the brand - people will now have heard of them and start paying attention to them. Like Transmeta hiring Torvalds.
The article is from the NY Times.
Do you even lift?
These aren't the 'roids you're looking for.
For the sake of easy readability, I'd like to give the grammar nazis somewhere to file all of their remarks.
do not read this line twice.
tweaks computers to let them tap vast amounts of storage at very quick rates
In other words, Yet Another Half-Baked Clustered/Distributed Filesystem we can add to the list of dozens of failed distributed/clustered filesystems.
Please help metamoderate.
It's a solid state hard drive that uses PCI Express instead of SATA. It looks pretty zippy. Decently large, too. Is there something else cool about their technology that I'm missing?
I had a very brief close encounter with him, in which I got to ride his Segway. He was, indeed, eminently approachable, with absolutely no "mightier than thou" attitude, self-assured, willing to engage, and very affirming to talk to.
(And that was *before* I recognized him!)
I've had the chance to play with some pretty phenomenal solid state drives (SSD) lately and, I have to say, that I can't believe that there isn't more industry buzz.
Depends on who you ask. The Sun ZFS guys are all over this and are screaming at the top of their lungs about the use of SSDs for both read and write performance:
http://blogs.sun.com/brendan/entry/l2arc_screenshots
http://blogs.sun.com/ahl/entry/hybrid_storage_pools_in_cacm
http://blogs.sun.com/main/tags/fishworks
Sun many have other problems, but engineering talent is not one of them.
Storage systems are not trivial pieces of hardware and the range of approaches for handling the problem is staggering.
In the red corner, you've your basic NAS and SAN solutions. In the blue corner, you've direct-disk-to-memory systems using RDMA and Infiniband. In the green corner, you've WAN solutions (SCSI-over-IP, RAID-over-IP).
In the purple corner, you've smarter drives (virtual sectors, filesystems in hardware). In the cyan corner, you've more powerful hardware (many read heads per platter, uber-large RAM caches).
(Knowing Wozniak's reputation for doing things different, he's probably inventing a rhododendron corner.)
There is no shortage of opportunity. However, as with the early home computer market, there is a shortage of consensus on what a storage system actually does, other than "store stuff". That seems to be a world Wozniak does well in - the lack of standards meant the Apple II did well, the presence of standards meant that NeXT didn't. In the current computing world, where standards are everything (especially if they come with pretty holographic stickers), can he do much with the flexibility in the arena?
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Want another penny? Here you go:
A cloud-distributed filesystem using each processor's bottom 2 or 3 general-purpose registers as a block for said filesystem, writing the contents only during certain times, or during periods of low access. This allows for lightning fast storage retrieval perfect for a database or large amounts of quickly updating information that needs to be retrieved just as fast, even better if archiving is not preferred after a brief period (think ticker tape), despite the possible redundancy of a RAID backup using said timing mentioned above. limiting factors are the speed of the reader(s), network speed, and bus bandwidth. Registers not used for storage are used for typical processing, aided by the amount of processors involved in cloud computing (think blue-gene).
There ya go, maybe I should make my own startup now?
PCI Express? Can you even boot from that without a specialized BIOS? Not unless this is intended for backup/secondary storage purposes, then its a good idea, especially if hot-swap/hot-spare safe, although most boards I've seen only have one (1) PCI-Express, so maybe there is much more missing than thought.
The "specialized BIOS" would be a ROM on the card itself - you can boot off of a PCIe SATA or SAS controller, just like you can boot off of a PCI PATA or SATA controller, just like you could boot off of an ISA ST506 or PATA controller.
(smacks self in head)
I actually set up a poweredge to net boot thanks to the ROM from a 3com card a looong time ago, I forgot about this was even possible because its been so long since I've had a need to do so (and virtually nobody tinkers with the BIOS that much anymore anyhow).
Woz doesn't have a corner, he sits in the middle of the room with a soldering iron, tinkering.
It would have an Option ROM, like RAID cards and every other bootable controller does
http://en.wikipedia.org/wiki/Option_ROM
Not using a SATA interface should yield a good performance advantage.
Rock on, Woz
If by "specialized BIOS", you mean "storage controller firmware on the card", then yes.
This is hardly different from any SCSI or SATA controller on the market, only this one has the "disk" built-in. When the system is POSTing, it triggers every device's initialization routine, which is where a disk controller can let the BIOS know it has (bootable) disks up for grabs.
-Billco, Fnarg.com
Let me translate this for you...
These are "LAN Solutions"
"SCSI-over-IP" - iSCSI
"RAID-over-IP" - some volume manager sitting on top of iSCSI
"WAN Solutions":
WAFS (Wide Area File Services) from the likes of Cisco or Riverbed. They optimize CIFS/NFS protocols which are horrible over high latency links.
Infiniband... Dying... besides infiniband used SCSI over IB to a IB to FibreChannel gateway.
Don't forget tape and our friend FICON.
Where can he be flexible? In the past few years we've seen the adoption of:
-Virtual Tape Libraries (tho they've been in the mainframe world for ages)
-Deduplication in Hardware
-Encryption of Data at Rest (in the tape drive; and now in the disk drive)
We've got plenty of CPU power with multi core systems... what about using that for Compression? (Sorry StorageTek did that in the 80s on their Iceberg (aka IBM's RVA Subsystem).
I don't need more capacity, I need to be able to manage it easier.
If a NeXT had cost the same as the Apple II I'm sure things would be alot different. And... NeXT was the other Steve...
Wow, I think I got a spam offering one of those.
"Is your girlfriend into bestiality? Don't want to share? Well, we have just the thing for you! Women don't know it's not canine!"
In other words, Yet Another Half-Baked Clustered/Distributed Filesystem we can add to the list of dozens of failed distributed/clustered filesystems.
Um... not even close?
This isn't a clustered/distributed anything. It's also not "virtual".
It's a very real, very fast, local storage for very real computers - servers mostly, but if you've got a few grand to blow on an extreme gaming rig, why not go the extra bit to make your levels load faster?
Their quoted numbers are per PCIe X4 device >100,000 IOPS and >640MB/s both reading and writing, and they have independent benchmarks back that up. They're not kidding. The game has changed. This changes everything about how traditional workloads are configured, when you use a SAN vs local disk, how much throughput your apps can get, how many VMs you can run in a server... basically everything in the server world except where you store the data. You still want to store the data in the SAN for redundancy reasons.
Help stamp out iliturcy.
It's a slam dunk. A no brainer. Two years from now you're going to be explaining to some young kid fresh out of college that "this is how we do it now. Forget that stuff they taught you." Again.
Right before you tell her to get off your lawn.
Help stamp out iliturcy.
Not to dismiss what he's done, but for being the chief scientist in a storage startup, it seems like he is a bit underqualified compared to what the cutting edge of storage looks like nowadays.
It seems that it may be more likely they brought him in in order to impress investors, i.e. an investor is more likely to put money into something where they have a big name of an entrepeneur that's struck it big. And it doesn't get much bigger than Wozniak.
You must be new here.
No, I be new here!
I got some bad grammar
The startup is on SSDs only. Look up their products page. products
So, you're in the wrong ring, though the same building.
...yeah, it's the buzzword. It's the current growth area.
Let's consider what The Woz did for floppies Back In The Day. While the early floppy drives are to modern drives the way the Wright Brothers plane is comparable to the B2 Stealth Bomber.... the fact is, The Woz turned the industry on its head. While in one light his contributions can be viewed as an incremental improvement, in every other light, HOLY CRAP HE KICKED SO MUCH ASS when it came to primordial microcomputer disk controllers. He proved that the highest-tech, super-chip-count hyper-expensive controllers could be implemented with a handful of chips.
And he could - COULD! - do it again.
I'm totally behind some company - ANY company - throwing money at The Woz, betting on the off chance he gets another flash of insight and pushes storage technology 20 years further ahead in as many minutes.
Was Woz the Right Genius at the Right Time, or is he a straight-up Hacker's Hacker, who just needs the right operational conditions for his genius to manifest?
Time will tell.
We have these in our production servers right now. They really deliver. They seem to top out at around 60,000 IOPS with EXT3 (the 100K figure was with XFS) but I've hit close to 800MB/s on sequential transfers.
It would have an Option ROM, like RAID cards and every other bootable controller does
http://en.wikipedia.org/wiki/Option_ROM
Not using a SATA interface should yield a good performance advantage.
Rock on, Woz
You could have an option Rom, or you could just emulate AHCI (or even ATA) in hardware up to the point the OS loads a native driver, and switch to native mode after that.
Actually I sort of wonder if you couldn't implement an AHCI contoller which talks to flash directly. The bottleneck in SATA is the drive and the SATA bus, not the PCI Express AHCI controller. PCI-E x16 can manage 4,000 MB/s compared to SATA2's 300 MB/s. SATA2 has plenty of bandwidth for a hard disk, but it looks like it will become a bottleneck with an SSD with lots of flash chips running in parallel. In fact an 2.5 inch Intel extreme SSD manages 250MB/sec now, pretty close to the SATA limit. A PCI Express card covered in NAND flash aimed at enterprise servers could easily be more parallel than this.
AHCI is quite flexible (it has efficient NCQ for example) and is already supported by all current OSs and Bioses. There's no reason why you couldn't design a wide flash array on a PCI express card that looks like a fast drive behind an AHCI controller to software.
The upside to this is that there is no device driver and option Rom to develop/support.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
I'd love to see Woz and Andy Bechtolsheim working on a new generation Thumper.
Want another penny? Here you go:
A cloud-distributed filesystem using each processor's bottom 2 or 3 general-purpose registers as a block for said filesystem,
Wow, that's like 24 bytes of storage per core. How will anyone ever fill it up?
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
Just saw that the Woz is also on the advisory board for IOActive, the guys who showed that your (RF)ID may be cloned off you from yards away...
The Hacker's Guide To The Kernel: Don't panic()!
There is no shortage of opportunity. However, as with the early home computer market, there is a shortage of consensus on what a storage system actually does, other than "store stuff". That seems to be a world Wozniak does well in - the lack of standards meant the Apple II did well, the presence of standards meant that NeXT didn't. In the current computing world, where standards are everything (especially if they come with pretty holographic stickers), can he do much with the flexibility in the arena?
I always thought that the Apple ][ did well because it was cheap and versatile, and that NeXT failed because their machines were outlandishly expensive and proprietary.
-- If you try to fail and succeed, which have you done? - Uli's moose
There's nothing new under the sun. We used to do it with system RAM dedicated to the purpose, back in the '80s.
Back in the stone age, we used to do it with RAM in a drive box. And then with add-in cards that acted like disk but stored RAM. I bet you noticed that RAM costs a lot of money if you need 320 GB of it. For a brief moment so long ago that I forget the date, we did it with something called "bubble memory".
I also talked about this here two or three years ago, before this product was produced, so I look forward to providing some prior art to the inevitable patent troll discussion.
But that's not the point. This is a startup, and they're at a vulnerable cusp in their history. If you need this product I suggest you buy it before somebody discovers the motive and method to kill it. I can think of three motives and two methods offhand, so if I were you I'd get crackin'.
Help stamp out iliturcy.
Infiniband... Dying...
How do you figure? Infiniband is the absolute final word in minimizing cost per port/density and provides rdma and ultra low latency on crazy high bandwidth connections. There is a reason that companies like NetApp use infiniband for their clustering solutions ;good luck maintaining cache coherency between two or more nodes over something else. Check out how scalable informatics is using IB links on storage boxes that can do over 5k iops at 1500 MB/s
-*The above statement is printed entirely on recycled electrons*-
1500 MB/s? Hmmm. Infiniband allows you to direct one or more lines (up to 12 in any given direction) to a given target and permits switches (and routers, but that adds latency) in that arrangement, so individual packets can be switched from individual input lines to any number of devices.
Currently, Infiniband over PCIe 2.x supports 5 GB/s per line, although any given server is also limited to 5 GB/s. This would limit you to 12 servers and 36 of the storage boxes you mentioned, on a single storage-area network if you wanted to guarantee zero packet loss at maximum throughput with no throttling or added latency. (Infiniband latency, btw, is around 8 ms, using the run-of-the-mill chipsets.)
To get the same sort of bandwidth on Ethernet requires that you channel-bond five 10 Gigabit Ethernet cards together. This is possible, it has been done (in fact, they've gone as far as ten in the laboratory), but it isn't a supported configuration by industry, it's more than the cost of a single Infiniband card for the same bandwidth, you have twenty times the latency of Infiniband, you don't get RDMA (except on the few Ethernet RNICs that exist, which cost more than Infiniband cards, so you're now paying more than 5 times the cost for truly comparable performance) and switching is a lot more complex. It's also a lot less reliable when channel-bonding to that extent.
If you need bandwidths in excess of about 20 GB/s, Ethernet is simply too expensive. Infiniband becomes cheaper at that point, as well as faster. If you're using something like RAC Oracle over any sort of size of cluster, then you'd be insane to use anything other than Infiniband.
I won't say it's the best technology in the world, but it's most definitely not dying. Even Netcraft won't confirm that one. Sure, for the absolute performance freak, Dolphinics SCI cards push 2.5 ms latencies, but those're point-to-point, not switchable, so it's directly attached storage only. It also doesn't do RDMA. Frankly, I would expect to see SCI dying before Infiniband, simply because it's not as versatile and the latency reduction is going to be more than devoured by the added number of context switches between the kernel and userspace.
However, SCI is not dying. It's not doing great, but it's surviving quite nicely. So any doom-and-gloom forecasts for Infiniband seem to be doomed to stay gloomy as they're not likely to come to pass.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Silly rabbit, Rings are for Tokens! (Or is that Tolkeins? I forget.)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
to SSDs, and that is lifetime. It is my understanding that the memory cells they are using in SSDs today are rated in the neighborhood of 100,000 writes. I do know for sure that it is less than 1,000,000 writes.
Within reason (many millions of cycles), magnetic storage media does not care how many times a particular location is written. Flash memory does. So "good" flash memory uses a scheme whereby the cells that are written to are rotated.
100,000 cycles is not very many for a computer. I could easily write a routine that would read and write bits in such a way to wear out large portions of your SSD within a few days.
That is not good. And because of the rotation scheme, it is not often mentioned by the SSD manufacturers.
How much would I need to pay to work there?
It must be like to have Mozart hanging around with your band...
http://www.dieblinkenlights.com
Woz hasn't done anything interesting in the last thirty years. I recognize his contributions are significant but just because he joins some start-up nobody has ever heard of doesn't mean it's going to be the next Apple or something. I suppose techies need their celebrities too and /. is their tabloid.
TRANSMETA
Catalin Braescu
Ofaly.com
...
how is babby formed?
Sun many have other problems, but shilling talent is not one of them.
there. fixed that for you.
job even in this economy. Damn you Steve for making me feel worse. But then I dont have to compete with you for other jobs out there. How lucky for me.
The text of this is very similar to a NYTimes article currently available on their web site. It sounds like someone is not giving proper credit.
The limitation is still there, even with wear leveling, which was my point (in fact I had already mentioned wear leveling). I could, indeed, write a piece of code that would wear out parts of your SSD within days. (Of course it would pretty much have to be done deliberately to cause breakdown that quickly... but it is possible.)
The fact is that SSDs have severe limitations. It is true that under normal usage, most people might not run into these limitations for a very long time. But we didn't "figure it out", if by that you mean actually removing those limitations. Under other than normal use, they can indeed be broken in a short time.
because you weren't paying attention. I was not referring to "normal use". I was simply pointing out that there ARE limitations, and even rotation (or "wear leveling", which I guess is the more appropriate term) does not address those limitations completely. For example someone who used the disk heavily, with the disk almost full, could find its capacity being reduced rapidly.
SourceN at http://www.sourcen.com/ said Well all good things must come to an end. Company is loosing one of its crown jewels.
If you think what I was saying is FUD. It is real, and I can prove it.
First you should understand (as I mentioned elsewhere) that I was NOT referring to normal use. Right?
You want to send me an SSD of yours to test? I guarantee I can write a program that will wear portions of it out within a few days. And after that, it is downhill all the way. (With, as you say, continually reduced capacity. Did anybody claim otherwise?)
Put your money where your mouth is. Send me an external SSD (external because my development machine is a laptop), and I will send it back to you with memory regions completely worn out. No cheating... I will even include the code I used to do it and slap an open-source license on it. The longer I have it, the more capacity will wear out, on a more-or-less linear scale. Oh, yes, there is a way to do it all right, and it is pretty damned easy. Theoretically, it could even happen to someone unknowingly... though I have no doubt that situation would be pretty rare.
Wear-leveling is great... for normal use. What I was pointing out -- for a reason -- was that for other than normal use (which does exist), it is far from perfect. These things do have real limitations. You seem motivated to try to deny that, but it is true.
No, I don't work for any HD manufacturer. I am just a simple software engineer... who knows her stuff.
1500 MB/s is because it is constrained by the disks. Point being that I agree that IB is awesome.
-*The above statement is printed entirely on recycled electrons*-
Sorry if my reply wasn't clear - I certainly wasn't intending any criticism of your post. I certainly got the point that you think IB is awesome, and I definitely agree. And, yes, you're limited by the disks, which is why the truly insane and the physics labs (there's a difference?) stripe across multiple drives.
With Infiniband, you get near-linear scaling because of the outrageous bandwidths and the ability to independently dedicate lines. Very roughly, 3 drives in parallel give you 4500 MB/s, which is just under what Infiniband'll do per line and what PCIe 2.x can deliver or absorb.
(HyperTransport 3 is even faster, but Infiniband network speeds seem to be locked to PCIe bus speeds. I'm guessing it's because it's widely used. You don't hear so much about HT3 these days - I can't honestly remember the last time Slashdot even mentioned the technology.)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
hkpsnstore@gmail.com would love some spam. I'm starting a collection