Slashdot Mirror
IT At the LHC — Managing a Petabyte of Data Per Second
schliz writes "iTnews in Australia has published an interview with CERN's deputy head of IT, David Foster, who explains what last month's discovery of a 'particle consistent with the Higgs Boson' means for the organization's IT department, why it needs a second 'Tier Zero' data center, and how it is using grid computing and the cloud. Quoting: 'If you were to digitize all the information from a collision in a detector, it’s about a petabyte a second or a million gigabytes per second. There is a lot of filtering of the data that occurs within the 25 nanoseconds between each bunch crossing (of protons). Each experiment operates their own trigger farm – each consisting of several thousand machines – that conduct real-time electronics within the LHC. These trigger farms decide, for example, was this set of collisions interesting? Do I keep this data or not? The non-interesting event data is discarded, the interesting events go through a second filter or trigger farm of a few thousand more computers, also on-site at the experiment. [These computers] have a bit more time to do some initial reconstruction – looking at the data to decide if it’s interesting. Out of all of this comes a data stream of some few hundred megabytes to 1Gb per second that actually gets recorded in the CERN data center, the facility we call "Tier Zero."'"
We need backup on floppy disk.
They may also be using something called load balancing, but we're still waiting for sources to confirm.
My wife, a staff physicist at FermiLab in their computing division, manages to keep me humble when I talk about the "big data" work I'm doing in my commercial engineering position. I think having to deal with a billion or so data points per day is big... Not so much in her universe!
I tried using the GRID - it's deeply embedded in acronyms and crud, practically impossible to use without a PhD. For crying out loud, it's just a batch farm!
Not yet.
Large Hadron Collider - powered by Linux
I was looking up how complicated the detectors were, and they were. They have 75M directional sensors and 9K energy detectors (calorimeters), each which are analyzed 40M times a second for "interesting" events. One out of a billion maybe recorded for subsequent deep analysis.
What do you want to imply?
That, somehow, he who does not know how to debug the kernel should not play with bit operations?
Something like that?
Or, that we should stop researching the structure of the universe, and instead focus on what we usually do, which is making war, screwing other people and post photos of our dicks on teh internet?
The head researcher will STILL come to IT and ask them to please help him sync his outlook contacts to his phone.
So they just used grep
Aah I see they use VMWare to manage the virtual machines. .. I guess Citrix is still lagging behind in the server virtualization field.
... if you know what I mean . :P
I am from Citrix
Roughly, assuming you can round it off to 53 weeks/year, if you do 1Petabyte/ear, and transferred that much constantly, that would be roughly 2887200000000000000000000000000000000000000000000000 BITS [individual 1s or 0s] per year
If you believe in privacy, and believe you have "nothing to hide" at the same time, you're a goddammed idiot
Did a bunch of work with some stock exchanges a few years back. It was an interesting environment and I see that CERN had the same problems that the stock exchanges had. They even had the where the number one budgetary item wasn't cost but electric load.
You only had so much power physically available in the data centers next to the exchanges and server rooms inside them. Monetary cost was never an issue, but electric load was everything. It seems funny considering their load is strictly a science based load and not monetary, but their requirements and distribution remind me greatly of the exchanges.
Well, that one should be called a petadick.
Achille Talon
Hop!
VMWare is pretty widely recognized as the king of virtualization-- at least so long as you arent concerned with money. Its overhead is far far smaller than the others especially when dealing with huge numbers of connections, and it simply has more features than its competitors.
Of course, that assumes you're willing to pony up for vRAM entitlements and Enterprise Plus.
Which doesn't mean those features are implemented well.
Not so long ago, I built an automated QA platform on top of Qumranet's KVM. Partway through the project, my employer was bought by Dell, a VMware licensee. As such, we ended up putting software through automated testing on VMware, manual testing on Xen (legacy environment, pre-acquisition), and deployment to a mix of real hardware and VMware.
In terms of accurate hardware implementation, KVM kicked the crap out of what VMware (ESX) shipped with at the time. We had software break because VMware didn't implement some very common SCSI mode pages (which the real hardware and QEMU both did), we had software break because of funkiness in their PXE implementation, and we otherwise just plain had software *break*. I sometimes hit a bug in the QEMU layer KVM uses for hardware emulation, but when those happened, I could fix it myself half the time, and get good support from the dev team and mailing list otherwise. With VMware, I just had to wait and hope that they'd eventually get around to it in some future release.
"King of virtualization"? Bah.
News that matters? The human race is not even able to handle itself and it wants to play with atoms.
I assume you're using a computer to post that? Maybe own a cellphone....?
That makes you a hypocrite of the worst kind. Sorry, but there it is in black and white.
No sig today...
King of virtualization when it comes to things like "supports live migration of a VM's execution state and/or permenant storage", or "stability and speed of the networking layer".
I cant speak to KVM as my experience is limited to VMware, and some HyperV and XenServer testing. But just doing a check from RHEV's own fact sheet, there are a number of things that are missing that are quite useful:
*Storage live migration
*Hot add RAM, CPU
*Hot add NICs, disk (note that RHEV has it wrong-- this does not require anything more than the free hypervisor, NOT enterprise plus as they claim)
*Live VM Snapshots (not really clear how RHEV doesnt have this, even Virtualbox has this)
Those are all pretty core features-- to my mind, ESPECIALLY the disk and NIC hot add. There are a lot of times that it is an absolute blessing to be able to roll out a new VLAN on the network and to just hot-add a NIC to the firewall VM on that VLAN, and your network suffers no outage. With disk, its awfully nice to be able to add a USB disk to the VM without having to reboot the entire thing (again, how does HyperV and RHEV not have this?).
The King Joffrey of virtualization, perhaps.
For large sets, this will be our guide even unto death, for the LORD will work for each type of data it is applied to...
That takes time. Time vs Space tradeoff.
For large sets, this will be our guide even unto death, for the LORD will work for each type of data it is applied to...
RHEV is getting there, still lacking some features and still rough around the edges. For instance:
Snapshots are a sore point but I've heard the next version out around this December will fix it and give us a truly live backup option. Most of the other issues you mentioned (hot-add of resources) will hopefully be part of the new version. Dunno yet if that means "total fix" but I've got my fingers crossed. Hey, it's better than KVM under RHEL...
Those with further interest in the article may find this informative:
http://www.geant2.net/upload/pdf/LHC_networking_v1-9_NC.pdf
Apparently, CERN uses BGP between T0 and T1, and uses only ACLs, no firewalls, for security.
http://www.enterpriseinnovation.net/content/brocade-delivers-100-gigabit-ethernet-solution-cern
just say'n
(oh come on, can you blame me? It freaking COOL!)
I can't speak to RHEV -- I ran on bare KVM. RHEV eliminated any feature from KVM that Red Hat didn't consider rock-solid enterprise-ready, effectively paring it down to the featureset they were willing to thoroughly test on all supported hardware.
All four of the items you listed were present and usable in upstream KVM back when I was working with it years ago, except for storage migration (which had just been implemented and was still at a place where I wanted to see other folks using it before I wanted to trust it myself... plus, already had shared storage anyhow and didn't need it).
The summary says it's 100 MByte to 1 Gbit, which is confusing in itself. I think "a few hundred megabytes" is correct. It's impressive to run at that rate continuously with high reliability, but it's nothing compared to Youtube and probably Facebook. If you say a "tweet" takes up 200 bytes including overhead, that's 500 000 tweets per second at 100 MB/s, so maybe even Twitter has to deal with that rate. The requirement for redundancy is probably stricter for the LHC, they have at least triply redundant storage for the original data. The data processing is maybe more demanding for LHC than for big internet companies, as all data are essentially equal, so to do an analysis, all data files have to be accessed (there is more filtering after the trigger, and intermediate files are stored containing reconstructed information).
Unfortunately that isn't saying much.
HAND.
Actually you're off by 26 orders of magnitude
1PB/s = 8e15 bits/s
8e15 bits/s *(3600s/h) *(24h/day)*(~365.25 days/year) ~= 2.5e23 bits/year
or 252,460,800,000,000,000,000,000 if you prefer counting zeros
even in stereo it'd only be 5e23.
--- Most topics have many sides worth arguing, allow me to take one opposite you.
ESX's cost is a bit of a PITA-- theres essentials plus, but of course that lacks DRS; and theres the free version which truly is nice for a single-server solution... but there are a lot of good contenders out there for less.
Im not gonna say that the others are garbage; I took a peek at Xen and really like that they dont gouge you to death for basic things like "can manage several servers at once". Im just saying that from my experience, as well as from listening to others in the recent ArsTechnica discussion on virtualization, it really sounds like VMware is still on top-- as long as you can pay.
Disclaimer: I just got my VCP and regularly do a lot of work with ESXi.
I'd like to know what their infrastructure looks like for storing that 1GB/s.
I was at OpenWorld in 2003 and they had some guy there from CERN giving a talk about how they were using Oracle9i (I read later that they upgraded to 10g, but no-doubt they upgrade to later versions relatively quickly), and he did mention that petabyte/s buzzword. It would be very interesting to know how it was all implemented, and how they manage to write 1GB/s to disk. Must be some serious RAC clustering going on, and some serious disk bandwidth capacity, but I wonder when they purge the data. After all, 1GB/s is going to require massive storage, and if you've gone with one of the big boys for storage like NetApp then it won't only be their research that could be termed as 'astronomical'.
"Everyone knows that vi vi vi is the number of the beast" -- Richard Stallman