Jim Handy of Objective Analysis writes:
I am conducting interviews of Fusion-io users and would like to speak to as many as I can.
If you use a Fusion-io product or if you know someone who does please let me know.
Thanks,
Jim
_______________________________
Jim Handy
OBJECTIVE ANALYSIS
Semiconductor Market Research
www.Objective-Analysis.com
PO Box 440
Los Gatos, CA 95031-0440
USA
+1 (408) 356-2549
Call Jim or write him.
The StorageMojo take
Jim is a friend and a conscientious analyst. If you have experience with Fusion-io products please give him a call.
And you can comment here too.
Courteous comments welcome, of course. I’ll pass along any interesting nuggets I get from Jim.
An angle which interests me is if anyone has tried using a in – memory DB, which could recover errors, taking a fast flash card as a target.
My interest is, though they are readily available now, getting the kit together to put a TB of RAM in a machine is a bit expensive for some near – line anaytics.
(Mind you, since it’s not an on-line app, i might have another go at that with some consumer spec cards soon, for fun, and not to be sniffed at savings)
IBM took a leaf out of Larry’s “I’ve no problem with open source, because i can take their stuff any time” angle, buying SolidDB, so I don’t know where that’s at. Project aborted because flash was really only just making noises at the time.
Sorry I can’t help with a lead. I spoke with quite a few people who were in touch with Fusion-IO when they announced, but the delay to distribution missed budgets. I think they OEM for HP now, so that confuses things, and perf data might be bound by some bigger support contracts.
Lordy, just checked their website, and they’re quoting some screaming numbers for their “Octal” product, but only for big blocks, which makes one wonder.
More importantly, it does look they are OEM only, Dell, IBM also.
Has anyone been able to purchase direct? The math goes funny with me once I’m messing with procurement, same for big – RAM machines, same for obscure flash array cards.
all best to all,
– j
Oh,
p.s.
sorry,
clarifiation: my project was aborted, but i put that in the context of SolidDB. Not to be confused!
also, a quick thought:
Quoting Woz on their home page worries me. That’s unadulterated marketing fluff. I think that’s weak, even a bit desperate.
Can no-one in Storage get a decent marketing guy any more?
If they all go on like this, storage will come aross as the dirty secret of IT, and people who have real budgets will find workarounds. (Who wouldn’t rather pay for the talent to write the code?) I guess EMC staff will be okay with the layoffs though, since they can sell snow to the eskimos 🙂
– john
In-memory databases probably won’t work with Fusion-io since it’s a block device.
You can buy through a distributor instead of the OEMs, but the price is the same.
Thank you Wes, for putting me straight. I probably should have said that we aborted before spending any time looking into hybrid in-mem stuff, but we got about nowhere, before line of business needs distracted.
By way of apology for mangling my posts above, i think I might have come across a user lead for the survey. Speculative, but there’s a kind of significant look from the developer at a slide with a Fusion-IO card, in the below linked presentation.
There’s also a wonderfully wry comment in this presentation, from the same dev lead, noting how he can’t believe how little architectural history devs he meets know. Something along the lines of “we just did what VMS did with clusters”, only inside the Intel micro-arch!
http://www.infoq.com/presentations/LMAX
Check it out, seriously enjoyable stuff. It got me thinking a whole different way about how much RAM might matter.
– john
Hi John,
I had the chance to test out the 80 GB Fusion IO drive with a previous company. Though I can’t post details, it did live up to the hype for performance numbers – though only through benchmarking utilities.
MySQL 5.X (0-5) – could not scale to utilize more then 10% of the offered IO. Greater performance came putting the temp tables (MyISAM) and binary logs onto the card rather then data – though we couldn’t achieve anything greater then a 40% performance gain.
I am not surprised vendors like 3par and Isilon have gone with cheaper Intel SSD’s then FusionIO. Distributing the I/O over a few dozen SSD’s can match the performance and take out the SPoF – though the technology is sweet and fast.
I also had to go through Dell or HP to purchase. We had passed it for Intel SSD’s as MySQL really could not utilize the I/O and cost didn’t justify the performance gain.
Thank you Anon!
Love it that someone still cares for ISAM 🙂
Your point about apps being unable to utilize bandwidth is very interesting, and why I’ll spend some “tinkering” budget on a far cheaper drive aimed at the gaming market.
I’m lining up to watch that presentation i linked for the third time – they were there to recruit, so left gaps. Like how to do stride lookups efficiently, keep the pipeline clean and full.
That’s your probem right there, ISAM was gloriously optimised to spindle speeds 🙂
Basically, with the new storage tech, we have to re-write code, or die.
I can point to a shed load of $MLN a piece kit about here which has no right to live, even before it was born. Mercantilism usually wins against Rennaisance thinking, but critically, when Rennaissane wins, it wins big.
There’s also some excellent criticism of the LISP style of coding in that presentation, and i like LISP a lot. Useless unless you know how it compiles in reality.
Basically, we have a embarassment of riches. The only way to deal with it is to get very deep, very serious. Cool. Sounds like why i got into this game.
“Cool Kit, Looking For Good Owner”
When you passed the FIO drive, for Intel SSDs – I did also – was it because of protocol assumptions in higher code?
loved your response, hit my buttons!
– john
John,
We have a few of the fusionIO drives configured in a stripe, the 640’s come actually in two 320GB volumes. So I have 4x320GB in a raid0 stripe using raid md0 on linux.
We currently see about 1.3GB/sec writes, and over 3GB/sec reads per second. Now that being said we have a hard time finding workloads that can stress them, but they do perform as expected.
We are currently testing with pSQL and working on getting it up to speed, currently around 70k/transations a second using pgbench -s 1000.
Hi John!
i’m pulling log files around, production LOB can’t even get drivers for this stuff.* So we’re doing a kind of circular check, holding state, instead of rewriting the dbm. Then another loop, against a snapshot. Ugh! That throws a bit more at a system. I’m doing this right now off – line, with a bunch of SSD, and nearline WD 4 series drives, LSI controllers (they speak T10) and Oracle HARD tricked onto Suse RT, because the traces are just good. Basically, what i could get with consumer MLC and a blagged long term test for the Suse, but with fiddling about. Spent more on clock synch. I’m close to maxing every throughput at the wire line. This isn’t so much TPS, because those benches pull, what, 20K entries? That’s several orders below possible tables in extremis. Yeah, it’s mapping. Sounds like a crappy framework web app gone wrong, but i hope we’re not being so silly!
The problem is that sql is a handy human abstraction, yet things like pSQL allow for direct tuples, which are a sometimes handy data abstraction. Knock out and recompile. If you map some trees, and follow your stored state, it’s usually possible to gain lots, for some cases, but that never helps the business end of things. Those are the guys who should work out how to do UI that works with this level! SQL itself is what optical engineers all “metameric” i.e. it only transmits part of the spectra well, so you look under a different light, and get different colors. No worries, i guess, because the opto science hasn’t fixed Macaulay Ellipses yet, hence why your digital camera – even those i bought on what i thought was the cusp for the price of a OK auto, and tried again, and keep trying, suck, badly, at color. (but are tremendously better than film) Mind you, you can still take great photos with any old thing, but you suffer for your art:)**
*I shall ditch a properly auditable DB over my proverbial dead body. Must eat. Some of the ideas, OK about 2 of 20, in that presentation, we’d thought about, but their best idea is to stuff the proper DB inside a closed loop with hard fails (actually, did they do hard fails? argument for raw network splits and time concurrency, but they didn’t put it that way, so they next thing is DMA over 10/e, think they hinted at that, but where are your signals, where is you cesium clock? can you rely on the embedded?) and tons of logging. Those streams add up. The neatest thing about what LMAX claim they did, is to internalize all this to uArch.
** Being more fundamental, your brain susses colors to normalize them at cumulative data rates yet to be worked out. Why you see “well” under sodium vapor on the road, e.g.
This isn’t my usual pitch, bcause i came close to it in tinkering only in a week during the past six months:
“tNTFS”
Transactional NTFS.
That was what was i think originally “Cairo”. Byte mag, ca. ’95 review.
It may not have T10, no nice checkums or volumes like ZFS does, but MSFT held out on their plan to stuf a kind of DB in NT. But NT drivers are quite nicely modular, so i can imagine a simple T10 or otherwise filter driver there.
I’m just looking into it now, again, not a casual thing to take a look, but if so – even for a 10% of “if so” – i rate their tenacity.
But, yeah, you get acid-“like” confirms on simple copies. Wonder how fast this is . . .
– j
Well, for what it’s worth, FIO just hired the best SE in the region for their Pacific NorthWest territory…
The 320GB model will be MLC based with similar performance. A few percentage points slower than the 80GB and 160GB, but still excellent. Any SSD manufacturer would steer a customer toward SLC products for servers. MLC has traditionally be slower and not as long-lasting, so it is interesting to see that Fusion-IO’s technology has overcome this according to their documentation. You are just buying the drive without worrying about which nand type is involved.
I want to thank Robin and the many folks who responded and helped with the survey. I got more respondents through StorageMojo than from any other source. StorageMojo Rocks!
Thanks, too, for the comments above from John, other John, Wes Felter, Billy, Anon, and Cialis (!!!!!) I have heard a lot of comments similar to yours about software not being ready to tap into this level of performance. Other John calls it an “Embarrassment of riches,” and that hits the nail on the head. I have been in touch with researchers at UCSD who are focusing a lot of attention on ways to optimize software or SSDs.
Just to be clear, the survey was not for Fusion-io, but for a start-up that wishes to remain anonymous whose intention is to make a companion product for the ioDrive and a growing number of competing PCIe SSDs.
Once again, a big: “Thanks!” to all for the help.
Jim,
thanks for your call out, but i added very little.
I have my full flame suit on when it comes to some things like, er, this:
http://venturebeat.com/2011/02/17/graham-y-combinator-no-tech-bubble/
but that’s because those people are simply not interested in low level arch. (that really i being polite) Tick tock. Me? never saw a bubble before, honest. 🙂
I get a early – production – not review – ship of 3rd gen Intel SSD any day now. (no hurry)
My contenion is this:
I prefer to do garbage collection where i can see it.
Did you see that LMAX vid? Did they need garbage collection? Did they not have massively irregular data patterns in their stream? Did the LSE not go down today, NYSE colo the other week? No good blaming national security. Write it properly. Grrr 🙂
back to the GC/storage:
Where is my Meta key?
How can i dump the whole machine trace, and come back?
But check out the latest vpro from Intel, that KVM is doing something similar i think on their boards.
The way of using this flash stuff, Wes put me onto the case. (again thanks)
Now, how do we replicate 30 and some years of doing good controllers, companies rising, busting, good tech passing on, with this new rather bipolar (huh, sorry) kit?
Really appreciate our words. But we ain’t out no woods. Bears be, ahem, here.
Robin, thanks for clearing up my messy double post the other day.
best to all!,
– j
This i think belongs here:
http://blog.serverfault.com/post/our-storage-decision/
hmm, Fusion I/O versus RAID10 Intel X-25’s.
Sorry i’m almost a month out of date finding that.
Anandtech benched the Intel 510s.