Episode 409 - HappyHakoween: Password Cracking Clusters, Remote Control Services

[chuckles]

Matt shows us how to turn anything into a service and provide a web frontend to manage them windows server, great for game server administration. Chris Gerling wraps up his three part series on Packet Sniffing with Wireshark techniques for packet filtering. Darren harnesses the CPU power of the HakHouse for good or evil to demonstrate cluster computing. Plus details on our Hak5 Halloween LAN Party!

Watch or download here. (http://revision3.com/hak5/HappyHakoween/)

-chuckles-

[bani-banan]

Audio out of sync very much?

Great episode.

Oddly, I was just thinking about building a cluster. I don't know why, I just want.

Weird.

[hak5matt]

Yes, we do acknowledge that there are some audio issues with this release, and we'll be working our hardest to get these ironed out for the next shoot, however we're still looking to get the final piece of our HD setup (mac pro), and if you can help out please visit http://www.hak5.org/stickers and donate.

Thanks for watching!
Matt

[rsone]

You can't use a cluster for f@h. A gigabit network can't handle the bandwidth between the nodes. Maybe with a Low-Latency Optical Network.

[bani-banan]

Yes, we do acknowledge that there are some audio issues with this release, and we'll be working our hardest to get these ironed out for the next shoot, however we're still looking to get the final piece of our HD setup (mac pro), and if you can help out please visit http://www.hak5.org/stickers and donate.

Thanks for watching!
Matt

I would buy a sticker, but the thing is - I live in Sweden.

The shipping alone would cost around $10-15.

Just post your paypal-address and I will send you a couple of dollars next week :)

[gimpbully]

folding@home uses a distributed model. It completely embraces the highest of latency networks. Honestly, you wouldn't need a cluster for it at all, you'll only introduce massive bottlenecks.
Assuming you guys can't afford a real interconnect(myrinet, infiniband, federation), you're going to want to be quite mindful of local memory bandwidth, to make up for a poor interconnect. Tasks like video encoding are great for these setups, you can feed mencoder binaries on each node a seperate offset of the input and just chug away. No massive collective calls, latency won't bite you as most of the task is localized to the node. You'll still have (inefficient) serial components of the workflow (joining chunks, muxing sound), but that's the nature of non-MPP workflows.

Software wise, stick with proven packages. If you're doing mpi, get mpich2, if you're doing smp, stick to openmosix. Cluster distributions are a dime a dozen, if you've got disk to spare in each node, just get Rocks and setup cfengine. Get yourself a small NFS server with a halfway decent raid (even software based is fine) setup. You should be getting better than single platter speed here, a 4 disk raid5 on even a sempron will get you 75+MB/s, respectable for a 5 node cluster. If you're feeling wild, get two gigabit switches -- dedicate one to NFS or whatever storage technology you're using and the other to the rest of the job traffic (mpi, etc). If you're *really* going to build a cluster, get yourself Maui/Torque and have real job scheduling (but really, you're just going to end up punching yourself in the face rather than deal with the annoyance of batch queues).

Above all, balance your hardware. Focus on getting the most memory bandwidth (ever wonder why opterons are still wildly popular in the HPC sector? on-die memory controllers and hypertransport), don't overload one node and underload another, it's only going to lead to idle cycles on the beefier machine. Get good NICs, make sure they do real checksum offloading (any processing you can take away from the proc, the better).

Remember, there are no shortage of resources out there regarding linux on commodity clusters. Just keep lookin around.

You can't use a cluster for f@h. A gigabit network can't handle the bandwidth between the nodes. Maybe with a Low-Latency Optical Network.

[rsone]

folding@home uses a distributed model. It completely embraces the highest of latency networks. Honestly, you wouldn't need a cluster for it at all, you'll only introduce massive bottlenecks.
I mean, you can't use a cluster to run the f@h smp client(cluster within cluster).

[gimpbully]

oh, absolutely, nowhere near all SMP code really benefits at all from an openmosix setup.
I mean, you can't use a cluster to run the f@h smp client(cluster within cluster).

[radzack]

Fun ep :)

[tuvedo]

where do i find KAOS to download

[davmoo]

Good episode.

My only disappointment was Snubsie only did one segment in those cute bunny ears. Well, that, and the part where I laughed so hard I sprayed iced tea all over the TV set during the show opening when you announced the last sponsor...gotta learn to not take a drink until after the show officially starts.

[computoman]

I agree, it was cool!.

[sembazuru]

For your Beowulf cluster, why not try to incorporate some PS3's into it. You can find them cheap (and hopefully working) on ebay, or get a couple new ones for $400/ea at your store of choice. Hak5 viewers can even donate their old PS3 when they upgrade (like send you their 20GB version because they decide they actually need the on-board WiFi for the upcoming AdHocParty feature). There are a couple academic clusters that I know of that are PS3 based being used instead of buying time on super computers. (~$3000 (8*$400 + extras) for unlimited use instead of ~$5,000 per supercomputer run.) Here are some links with explanations:

PS3 Gravity Grid at UMass@Dartmoth being used for gravity wave modeling calculations. Started as an 8 PS3 cluster, currently a 16 PS3 cluster. I found a citation on the web claiming that in it's original configuration of 8 PS3s it had comparable computing power to 200 Blue Gene supercomputing nodes. Unfortunately, this site is quite sparse on the details:
http://gravity.phy.umassd.edu/ps3.html
Read this Wired article for an amusing story on how the first 8 PS3's were obtained:
http://www.wired.com/techbiz/it/news/2007/10/ps3_supercomputer

NC State University has what is considered the first PS3 cluster. (I haven't done any date comparisons myself, so I don't truly know who was (on) first.) This is their press release, sparse on details:
http://moss.csc.ncsu.edu/~mueller/cluster/ps3/coe.html
But, this page goes more into depth of what the systems specs are, what OS and software is being used, etc. Looks like the target audience for this page is students at NCSU wanting to use the cluster:
http://moss.csc.ncsu.edu/~mueller/cluster/ps3/

Here is a PDF of a power point given for a symposium at University of Alaska Fairbanks that goes into some of the hows and whys. (Probably better as a backdrop to a talk instead of what we get by just looking at the slides.) There are some interesting bench marking slides near the end:
http://www.arsc.edu/science/multicore/Nosum_PS3.pdf

Here is what you can buy from Terra Soft, useful for seeing what software they use. They use a non PS3 computer as the head node, probably because of the 256MB RAM limitation of the PS3... Not that I'm suggesting plunking down the +$17K Terra Soft is asking (that's a lot of stickers), but more of a resource to see how others are doing it.
http://www.terrasoftsolutions.com/store/index.php?submit=software&submitimg[hardware][solutions]=1

Now, after all that, I'm not sure how well PS3s in your cluster would work with what you want to crunch on the cluster. But it might be interesting to see how adding 1 or 2 PS3s from eBay to your cluster changes the performance. What tasks are greatly enhanced by have the PS3s there, what tasks work the same (or worse) with the PS3s there. Etc.

[bani-banan]

^ Copied and pasted that onto a document for future need.

Thanks for the info:)