Slashdot Mirror
Storing CERN's Search for God (Particles)
Chris Lindquist writes "Think your storage headaches are big? When it goes live in 2008, CERN's ALICE experiment will use 500 optical fiber links to feed particle collision data to hundreds of PCs at a rate of 1GB/second, every second, for a month. 'During this one month, we need a huge disk buffer,' says Pierre Vande Vyvre, CERN's project leader for data acquisition. One might call that an understatement. CIO.com's story has more details about the project and the SAN tasked with catching the flood of data."
From my experience, generic blue work clothes (preferably with your name on the breast pocket) work best. I once got into some research facility (they had lasers and everything) because I got out of the elevator on the wrong floor and some guy in a lab coat opened the door for me (I was wearing my work clothes because I was on my lunch break). I wandered about at the place for something like 10 minutes before I found a way out. There was even a security guy of some type sitting at a hallway but he lost interest in me after I looked him in the eye and said hello.
We are expecting to record around 15PB / year during the LHC running. This data is stored onto magnetic tape with petabytes of disk cache to give reasonable performance. A grid of machines distributed worldwide analyses the data. More details are available on the CERN web site www.cern.ch.
load"*",8,1 would load something from a diskette, not a cassette.
It's only 5x HD SDI single channel ~ 200MB/s. Any major studio could handle this with ease.
SDI is how the movie guys move their digital stuff around. A higher end digital camera will capture at 2x HD SDI for a 2K res, 4:4:4 colour space. A few of em' and you got your 1GB/s easy. Spools onto godlike RAID arrays.
Get em' to call up Warner Bros if they have problems.
Right now, the average event size for ATLAS is 1.6 MByte and the system is designed to keep around 200 events per second, or roughly 300 MByte. This isn't much of course, but you have to consider that the bunch crossing rate (i.e. the rate at which bunches of protons will collide and generate events) is 40 MHz.
So you have to design a system that boils this rate from 40 MHz down to 200 Hz and only keeps the interesting parts, while also buffering all the data in the meantime. For this reason, the first trigger level is entirely implemented in hardware right in the detector and reduces the rate down to 75 KHz with a latency of 2.5 s. The rest of the trigger works on clusters using Linux computers and has a latency of o(1s).
- Printed hardcopy. Many authorities recommend this as you do not need to worry about changes in data formats over time. For exact calculation, we would need to know the font they were planning to use and the character encoding. However, let's take a working assumption that they can cram 10KB of data onto an A4 sheet. That implies 259,200,000,000,000 pages. They will probably not want to use an inkjet printer if they use this solution and may, indeed, choose to acquire multiple printers and split the load. A single printer at 10 ppm would take approximately 50,000 years to complete the backup. On 70gm paper, it would weigh a little over two million tons. At any rate, this would certainly produce reams of output.
- Diskettes. This was good enough for nearly everyone 15 years ago. It is curious that such a tried and trusted technique is no longer in fashion. I assume regular 3.5" 1.44MB diskettes, generally recognised as easier to handle than 5.25". We shall need around 1,800,000,000 diskettes. One drawback is the person changing the diskettes as each one filled up might become a little bored after a while. On the positive side, the backup will be quite a lot faster than the printed solution. Assuming about one diskette per minute, inclusive of changing disks, the backup could be complete in less than 3,500 years.
- Now considered somewhat old fashioned, punch cards were once a mainstay of every programmer's personal backups. Like printed hardcopy, anyone familiar with the character encoding used, could read the data without needing any access to a computer. If we assume 80 column cards, we would need 32,400,000,000,000 cards. I would be somewhat concerned about the problem of getting this stack of cards back in the correct order if I dropped it. With a weight of about 30 million tons and stretching perhaps 6 million miles end to end, handling certainly would be challenging and an accident very possible.
- Paper (punched) tape was the only alternative on the first computer I used, a basic early model Elliott 803 without the optional magnetic tape. If I recall correctly, you could manage about 10 characters per inch, so you would need a paper tape over 4,000,000,000 miles long. Hmmm, that would be silly. The other solutions are clearly better.
I am sure other options will be considered, but I just wanted to bring these up in case CERN had failed to consider themYou may think of it as product placement, but I use it. I even provide the occasional blog entry on it on Advanced Topics. I sat through a RedHat performance tuning class that was quite excellent. But when they came to the part about ext3 and tuning it, well, let's face it - ext3 just isn't going to scale. I started with Veritas' Filesystem which is pretty nice. If you're a small-time admin, then you never get beyond a local, 4U disk array. Once your group spends more than US$2million on servers though, it's obvious what the problem is: Storage - The Final Frontier. SAN and clustered filesystems allow a level of scalability completely unheard of before.
They also completely left out anything but a tagline of their multi-tiered solution. I wish they'd talked more about how CERN supports 500Gbit per second aggregate throughput to their disks (at least they implied that). 50GB/sec (or so) is probably the toughest I/O problem you've ever dealt with, or will deal with for a long time. Whose RAID controllers did they use? Did they focus on speed (ASIC and ISL minimization), availability (redundant fabrics), or both? Did each node get dual 4Gb links or just one?
If this had been an advertisement, they would have discussed some 3.0 features like LAN clients.
So, in short, it's easy to say it sounds like an advertisement. Quite possibly, Quantum (formerly ADIC) coerced them into getting the piece written. But if this had been an advertisement, there is so much more that is going on under the hood that would have been said. Large, fast, distributed filesystems are non-trivial and take an extreme amount of engineering and testing. StorNext really is good at what they claim to do.
If you want to read about some of the drawbacks though, I yak about them on my blog. Sorry for the plug.