| === CyberJacob is now known as CyberJacob|Away | ||
| === mwhudson is now known as 5EXAAR2RG | ||
| === 5EXAAR2RG is now known as mwhudson_ | ||
| bigjools | jtv: one thing I forgot to say that we need to think about is to have a way indicate that particular nodes should use a particular bootresource | 03:51 |
|---|---|---|
| bigjools | tags seems the obvious choice, but it would need hooking up | 03:52 |
| jtv | Shouldn't that be controlled by the settings we already have, plus a "label" field? | 03:53 |
| bigjools | jtv: how would we currently direct a specific node to use a specific kernel? | 03:53 |
| bigjools | once labels are implemented, I mean | 03:54 |
| jtv | Aren't arch/subarch/release/label/purpose already the main constraints for that though? | 03:55 |
| jtv | Although I guess tags make sense for HWE. | 03:56 |
| === zchander_ is now known as zchander | ||
| === Lord_Set2 is now known as Lord_Set | ||
| Lord_Set | Greetings | 11:05 |
| jtv | Hi Lord_Set | 11:09 |
| Lord_Set | How are you jtv? | 11:11 |
| jtv | Fine here, thanks. You? | 11:11 |
| Lord_Set | Good. Ready for work to slow down some lol. | 11:12 |
| jtv | That's got to be good, the rate you've been going. | 11:12 |
| Lord_Set | 20+ hour days for the past few weeks... | 11:12 |
| jtv | /o\ | 11:12 |
| Lord_Set | Also have some pretty awesome new things in the works. Can you imagine 4800 cores, 1200 phsyical nodes, within 1 rack? | 11:14 |
| jtv | Nice rack. | 11:14 |
| Lord_Set | That will be a good test for MAAS ;) | 11:14 |
| jtv | Yes it will. | 11:14 |
| Lord_Set | We got the wild hair up our asses about building a scalable and very cost effective solution using ARM. | 11:15 |
| jtv | Only way you're going to get those numbers! | 11:16 |
| Lord_Set | Using quad-core A9 or A11 boards with atleast 8gb DDR3 and 8gb internal flash. | 11:16 |
| jtv | Would be interesting to see where the bottlenecks are. If you have a large number of those racks, you're going to need a fast server. | 11:16 |
| Lord_Set | Yeah | 11:16 |
| jtv | You might hit tftp bottlenecks, in which case... smaller clusters. | 11:17 |
| Lord_Set | Well they would all be running 1g ethernet... currently we are already doing testing and using dual 10g on our cluster controller | 11:17 |
| jtv | *slobber* | 11:17 |
| jtv | I bet YouTube's pretty good on that. | 11:18 |
| Lord_Set | We could easily do 2 10g dual port nics in a port channel | 11:18 |
| Lord_Set | So have 40g possible for the cluster controller... I am sure that could handle it all | 11:18 |
| jtv | We've seen some slowness (ages ago though) with the tftp server; it may not be a network bottleneck. | 11:19 |
| Lord_Set | Which tftp server package are you guys using for MAAS? I never really looked into it. | 11:19 |
| jtv | We got it for free with something else, but I need to remember what that was. | 11:22 |
| jtv | It's Twisted-based. | 11:22 |
| jtv | Haven't heard anything about slowness lately, so maybe that was just something that went away. | 11:22 |
| jtv | tx-tftp..? Lemme look. | 11:23 |
| Lord_Set | Hmm | 11:24 |
| jtv | Yes, python-txtftp. | 11:24 |
| Lord_Set | I have so ideas that you could could do to internally cluster or load balance TFTP to increase throughput... TFTP is a finicky and picky protocol. | 11:24 |
| Lord_Set | err have some | 11:25 |
| Lord_Set | I will shoot them to you privately in a bit | 11:28 |
| jtv | Let's hope we don't need to get too creative though! | 11:28 |
| jtv | Especially when it's third-party code. | 11:28 |
| Lord_Set | True. I could have one of the developers on my team write a new custom built multi-thread and instance tftp server. | 11:30 |
| Lord_Set | If we start to run into some issues... We will see soon as even outside of the ARM scenerio we will be hitting super high density between physical servers and VM. | 11:30 |
| jtv | The thought of getting that makes me a little giddy... But as always: remember Knuth's Law or suffer the consequences. :-) | 11:31 |
| Lord_Set | Should be hitting 160-200 MAAS nodes per rack depending on the specific hardware used. This will be deployed within a month max. | 11:31 |
| jtv | Are you expecting a big jump to the big racks, or a steady ramp? | 11:33 |
| Lord_Set | To the 160-200 node racks will be a big jump and move. | 11:33 |
| jtv | And to the 4K-node racks? | 11:34 |
| Lord_Set | The ARM idea is a development idea and most likely 3-6 months off. | 11:34 |
| jtv | Ah OK. | 11:34 |
| Lord_Set | We will have to have some hardware fabricated for it... a power and ethernet backbone for each rack case | 11:35 |
| jtv | Were you thinking to have an x86 cluster controller with ARM nodes? Or all ARM? | 11:35 |
| Lord_Set | x86 most likely just for the power. | 11:35 |
| jtv | Yeah. | 11:35 |
| Lord_Set | Also unsure on exactly what will be their best use. Possibly Hadoop or cheap web server nodes. | 11:36 |
| jtv | Also, Python is quite memory-hungry so with large numbers of objects, address space might become an issue if we're talking 32-bit ARM. | 11:36 |
| jtv | I don't *think* we're at the level yet where a few thousand nodes might be an issue, but the proof of the pudding is in the eating. | 11:36 |
| Lord_Set | Well I would prefer to use 64bit Intel Alterra ARM chips. | 11:37 |
| Lord_Set | Or something similar | 11:39 |
| jtv | Haven't looked at those. Are they available? | 11:39 |
| Lord_Set | Pretty sure they are | 11:39 |
| Lord_Set | Pretty awesome 14nm chips | 11:39 |
| Lord_Set | You can get anything ARM produced super cheap in China lol | 11:40 |
| jtv | I'm sure you can! | 11:40 |
| jtv | Or Taiwan, I guess. | 11:41 |
| jtv | Is the Altera licensed as a design, or only sold as a chip? | 11:41 |
| Lord_Set | How do you think companies like Minix can sell quad-core arm boxes with 4gb ram and 8gb flash, hdmi, optical audio, gigabit ethernet, 2x usb 2.0, 802.11n, a remote and cables and power for 100 dollars per... | 11:42 |
| Lord_Set | That includes packaging and shipping | 11:42 |
| Lord_Set | More or less the same chip that is in the Samsung Galaxy S4 | 11:43 |
| jtv | Yeah that world never ceases to amaze me. | 11:43 |
| Lord_Set | ARM is for sure the future of high density computing | 11:44 |
| === edu-afk_ is now known as edamato | ||
| tych0 | hi bigjools, no luck even with syncdb on running the maas dev env | 17:34 |
| tych0 | bigjools: now i'm seeing http://paste.ubuntu.com/7080407/ | 17:34 |
| tych0 | bigjools: do you guys have a 'just-run-the-damn-thing' target in the makefile? | 17:35 |
| === edamato is now known as edu-afk | ||
| === zchander is now known as zchander_ | ||
| bbcmicrocomputer | can you create a zone spanning multiple maas-clusters/subnets? Or do zones create groups within one cluster/subnet? | 20:41 |
| === kevin is now known as Guest68425 | ||
Generated by irclog2html.py 2.7 by Marius Gedminas - find it at mg.pov.lt!