/srv/irclogs.ubuntu.com/2026/09/21/#ubuntu.txt

DrMaxeelstrebor : not upgrading due to phasing is for packages that are for the next (sub)release but released now00:36
DrMaxsay, something that would be with 26.04.1 but you're still in 26.04(.0)00:37
DrMaxnot upgrading are usually packages that will require some other package to be removed00:37
DrMaxhttps://help.ubuntu.com/community/PhasedUpdates00:38
DrMaxpackages that are "not upgrading" can be upgraded with full-upgrade00:40
DrMaxit will warn you normally: upgrading X, removing Y00:40
DrMaxsudo apt full-upgrade00:41
DrMaxyou will get the usual behavior: upgrading X, removing Y, [y/n]00:41
eelstrebori kinda figured that i would have to force the upgrade of some packages01:12
=== rebel is now known as Guest9558
=== deepend1 is now known as deepend_tilde
=== Seb_ is now known as Sebastien
DrMaxsomewhat unrelated to ubuntu itself, but linux-centric: is there a framework that allows distributing processes across several linux boxes as if it was one big machine?04:31
DrMaxof course, there must be, I meant, which one is favored04:31
computelabs_inHello good morning fellow!04:36
DrMaxyoes04:37
=== Remy_ is now known as Remy
twobitspriteDrMax: generally it's considered best practice these days to just have a program on each syste that communicates over a high-speed network connection. sharing raw memory across machines is doable, but generally not optimal04:48
DrMaxthey don't really need to communicate... just enough to know on which part of the data to work04:52
twobitspriteDrMax: you want a dispatching system, not "the same process running on multiple physical computers"04:53
twobitspriteDrMax: check out the Python framework called Celery04:54
DrMaxI also need to know when they're dead04:54
DrMaxso I can launch some more04:54
twobitspriteDrMax: are you running this on AWS or something? Because you can have liveliness checks on nodes and auto-scaling.04:55
twobitspriteKubernetes also supports the same concept04:55
twobitsprite(AWS also supports Kubernetes with AKS)04:56
DrMaxno, at home. I have several machines.04:59
DrMaxwhat I want is run the same program on different chuncks of data, but making sure that if a core is free, I can launch another instance05:00
twobitspriteDrMax: there is no good way to just take an existing program and "just add processors" by linking up other computers. processes need memory, which means transfering memory over a network05:00
DrMax(different chunks of data or with the same but different parameters, same difference)05:00
=== timj2 is now known as timj
twobitspriteDrMax: you're going to have to do some programming. What's the program you're trying to run on multiple machines?05:01
DrMaxa compression algorithm. chunks are independent05:01
twobitspritewhat compression tool are you working with?05:02
DrMaxI'm writing it. Not bzip or something like that05:02
twobitspriteI guarantee that trying to transmit chunks over your CAT6 LAN is going to be slower than just running it on a single machine05:03
DrMaxbut the algorithm admits a wide number of possible internal parameters, and one combination will give the best results.05:03
DrMaxit's basically an hyperparameter search05:03
=== rebel is now known as Guest3239
DrMaxIt's compute-intensive, not io-intensive05:04
twobitsprite"a wide number of possible internal parameters" just screams "lots of network traffic" if you're trying to cluster this05:04
twobitspritecompute = memory05:04
twobitsprite= network traffic05:04
DrMaxone block, a large number of hyper-parameters.05:05
DrMaxNot IO-Bound.05:05
twobitspriteand how will your cluster nodes get those "hyper-parameters"?05:05
DrMaxthe parameters can be passed as arguments, from the launching script, I guess?05:06
twobitspriteare all of your nodes connected to a SAN or NAS with the same raw data to work with?05:07
DrMaxif you launch a process and tell it try all values from 10 to 1000, you don't need that much communication05:07
twobitspritedo you have a Fiberchannel switch?05:07
DrMaxhmmm I guess the files can be replicated locally?05:07
twobitspriteare you compressing data, or guessing numbers?05:07
DrMaxyes05:07
twobitspritelol05:08
DrMaxIt's like asking are you training a NN or are you guessing numbers?05:08
DrMaxWell, that's how hyperparameter search works05:08
twobitspriteright, but it still needs access to the raw data you're compressing05:09
DrMaxI have enough local storage to replicated the data set before hand05:09
DrMaxto replicate*05:09
twobitspriteare you just doing this once, or is regular compression performance a requirement?05:10
DrMaxotherwise I'd have to go with *shudders* NFS?05:10
twobitspriteare you just trying to copy your MP4 collection somewhere?05:10
twobitspritewhat's wrong with NFS?05:10
DrMaxActually it's research. I'm not trying something silly like fiddling with ffmpeg until my animes are small enough05:10
DrMaxI had difficulties in the past with it05:11
twobitspriteno offense, but you probably weren't using it right05:11
DrMaxprobably.05:11
twobitspritedo you know what jumbo frames are?05:12
DrMaxNone taken, I'm not a linux admin05:12
DrMaxI use it05:12
DrMaxyes, I know what jumbo frames are05:12
DrMaxI'm still only wired in Gbit, though, even if I have a couple of boxes with 2.5, 5 and 10 gbit ethernet05:13
twobitspritethat's going to be a huge bottleneck for large data xfers05:14
DrMaxFor a first exploratory phase, replicating a couple of tens of GB of data is not really a problem05:14
twobitspritewhy can't you run your compression algorithm on tens of GB of data over SATA and a few cores of CPU?05:15
DrMaxBecause the hyperparameter space is big.05:16
DrMaxIf I can divide time by 100, it'll show05:16
twobitspriteis it a memory constraint?05:16
DrMaxno, not really. It's really compute-bound05:16
twobitspriteyou can't just "divide time by 100" by adding a few ethernet-connected computers05:17
DrMaxWell, I have 200 cores available05:17
DrMaxthat should help05:17
twobitspritethe limit is the ethernet05:17
DrMaxit's not like I'm using my second laptop05:17
twobitspritebring in FiberChannel, at least, and then we'll talk05:18
DrMaxHmm not sure if true. I can slice the hyperparameter search across many cores, each core will work at least a few seconds05:19
DrMaxit's not communication or io-bound05:19
twobitspritewhat will the hyperparameter search work with?05:19
twobitspriteit has to work with real data, right?05:20
DrMaxthe working set is somewhat small, and the data is read sequentially05:20
DrMaxtrue05:20
DrMaxbut the data is local on each box05:20
DrMax(or, the data can be made local)05:20
twobitspritehow long does it take to make the data local?05:21
DrMaxonce? How long to copy 20GB?05:21
DrMax5 minutes max?05:22
twobitspritesounds like a performance testing exercise05:23
DrMaxI guess it can be seen like that05:23
twobitspriteis distributing the data over 5 minutes faster than running the algorithm locally?05:23
DrMaxno05:23
DrMaxThe goal is to gather lots of data about how the algorithm performs with the various hyperparameters05:24
DrMaxit's not meant to be practical just yet05:24
twobitspriteare you a doctoral student? because this is sounding increasingly academic05:25
DrMaxI am a tenured professor doing research05:25
DrMaxbut yes, it is very doctoral-like05:25
twobitspriteso, this is "lets make this worse to prove a concept"05:25
DrMaxworse in what way?05:26
twobitspriteslower05:26
DrMaxhow, "let's divide a very large expensive-by-instance hyperparameter space search across as many core as possible" slow?05:27
twobitspriteI thought we already established that...05:27
twobitspriteethernet05:28
DrMaxIT IS NOT IO BOUND05:28
twobitspriteyou just said it would about 5 minutes to copy the data over. I asked if that would be faster than running it locally. you said "no".05:28
DrMaxthat's a standard corpus05:29
DrMaxevery one uses it05:29
DrMaxIf I copy once, that's it. That's the setup for the experiment05:29
twobitspriteeveryone in your university? or everyone in real life?05:29
DrMaxthen each computer can cache it all it wants05:29
DrMaxSome disciplines do have standard benchmark data sets05:30
DrMaxIf you do medical image compression, it helps greatly to use the same xray image database05:30
twobitspritesure, the setup for the experiment... must be nice working in academia, using pre-cached data for your tests. what exactly do you think your experiment will prove in real-life scenarios?05:30
twobitspriteif the image is already compressed, what's the point in a clustered compression algorithm?05:31
DrMax...05:31
twobitspritethe entire point of data compression is to reduce transmission time05:31
DrMaxwhen you use an algorithm that has been standardized05:32
DrMaxbut you see, sometimes you want to develop new algorithms05:32
DrMaxexample of standard test-set: https://loki.disi.unitn.it/RAISE/05:32
twobitspritenew algorithms are great, but they have to be functionally better than previous algorithms05:32
DrMax...05:32
DrMaxok, that's getting nowhere05:33
twobitspritethe difference between what you're describing, and RAISE, is that each node would have to download each image anyway. But if the data you're trying to compress is on a local HDD on one node, there's no reason to transmit "chunks" of it to other nodes unless you have an extremely fast network05:35
DrMaxI'll try to give you a better idea of what the algorithm does (without entering in the details)05:36
DrMaxlet's say you have a function that takes a block of data and a few "tweak" parameters and returns a score (the compression achieved, or quality, or something else)05:37
DrMaxf(block,param1,param2,param3)05:37
DrMaxeach param_i are independant05:37
DrMaxeach are drawn from a large number of possible values05:38
DrMaxyou could do, naïvely, three nested loops and do that sequentially05:38
DrMaxpick the best one05:38
DrMaxand say param1=3, param2=281, etc.05:38
DrMaxthat's the best value, next block05:38
DrMax...but that's silly to do sequentially05:39
DrMaxyou should do that with as much parallelism as possible05:39
DrMax(by block, by file, etc.)05:39
DrMaxbut once a given block is loaded in ram, you'd use shared memory and as many parallel threads as possible to search the hyperparameter space for that block05:40
DrMax(while, arguably, some other computer does another block, or some other region of the parameter space)05:41
DrMaxIf the future, maybe a GPU-based implementation would allow more parallelism, I don't know05:41
DrMaxFirst step is to evaluate if the new algorithm does, or can, at least in principle, do better than the previous, similar algorithms05:42
DrMax(compression-wise, I mean)05:42
twobitspriteI wish you the best of luck05:44
DrMaxSooo... the initial plan is to launch, on possibly a large number of different computers, searches on limited regions of the hyperparameter space05:44
DrMaxjust like what you'd do with a neural network05:44
DrMaxyou train networks independently with different initialisation+hyperparameters05:45
DrMaxyou don't do that sequentially if you have a large number of cores available05:45
DrMax(I guess you'd say you'd use a large number of GPUs...)05:45
twobitspriteI just think you're only CPU-bound until you're suddenly IO- or network-bound. If you don't have high-end networking (FiberChannel, iSCSI over 10BG with jumbo frames, etc), I suspect you're going to end up IO-bound.05:48
mujaasiHey I am new here i dowloaded mint05:49
DrMaxMaybe05:49
DrMaxYou were speaking of "dispatch system" earlier05:49
DrMaxwhat did you had in mind?05:49
DrMaxmujaasi : welcome05:50
twobitspriteDrMax: maybe if your algorithm is in-memory only (i.e., the client nodes aren't also SATA-bound) and are serving their chunk directly over high-speed ethernet or FiberChannel....05:50
mujaasithanks Drmax05:50
twobitspriteDrMax: Celery is an HPC library for Python05:50
DrMaxit is mostly in-memory, I suspect. The blocks would be much larger than ... 1GB I guess ?05:51
DrMaxHmm I could have a python think launching local instances05:51
DrMaxCelery is a task queue thing... (what I get from the blurb)... that might be the thing05:52
twobitspriteDrMax: I've been a sysadmin/DevOps for 20 years, and I've seen people try hairbrained clustering solutions over and over again... it's always network- or IO- bound, even if the initial system seemed CPU-bound05:52
twobitspriteDrMax: I genuinely wish you the best of luck, but I just want to temper your expectations05:53
DrMaxEach individual machine is rather capable (32, 64 cores, with 128GB ram)05:53
DrMaxI expected I could cache data in ram in each05:54
DrMaxby far the easiest solution is to use a single machine with OpenMP05:54
twobitspriteDrMax: yeah, I'm not worried about the CPUs. It's one of those "9 women can't produce a baby in 1 month" kind of scenarios. Just throwing more CPUs at is isn't always the solution.05:54
DrMaxI dont think the analogy holds, but ok, I understand what you're saying05:55
twobitspriteDrMax: obviously I don't have a full flowchart of your algorithm, or how you're providing the data to each node. I'm just telling you my experience with HPC.05:56
DrMaxit's just at the PoC level, of course05:56
DrMaxit may be shit, erm, "unsatisfactory" anyway05:57
DrMaxbut it's, I think, in the "embarassingly parallel" category05:58
twobitspriteIf each node was connected via 10GbE iSCSI with jumbo frames, or FiberChannel, to a multi-platter disk array, like a NetApp or EMC, *maybe* clustering the algorithm would improve performance05:58
DrMaxHow much faster is FiberChannel?05:58
twobitspriteup to about ~120 Gb/s, but it's not even the bitrate, it doesn't have the protocol overhead of TCP/IP, and the back-and-forth SYN/ACK stuff06:00
twobitspriteit's not ethernet, so it has it's own network switches that speak a totally different protocol because it doesn't have to route internet traffic06:01
DrMaxand how is it supported by the os?06:01
DrMaxand the applications?06:02
twobitspritethere is hardware like PCIe cards, then OS drivers, then you mount a LUN (Logical UNit) on the OS, and the application just sees a filesystem06:02
DrMaxI'm reading on that06:05
twobitspriteBrocade and Juniper were FC switch providers I've worked with the most06:06
twobitspritethey're pretty solid, but I've also heard good things about Cisco's FC switches. (obviously Cisco is king of ethernet switches, but FC is mostly Brocade's domain)06:08
DrMaxand how expensive is it?06:08
twobitspritevery06:08
DrMaxvery 2000$ or very 200k$ ?06:08
twobitspriteI don't know. it's one of those things the accountants and managers work out without inviting me to the meeting. like a restaurant that doesn't print the price on the menus06:09
DrMaxI see06:10
DrMaxUnfortunately, I have the bill and it's easier to get a research grant if it's not *that* expensive06:10
twobitspriteI worked as a level 3 escalation engineer at NetApp for 3 years and I never saw a billing price on anything06:11
twobitspriteDrMax: iSCSI is the next best if you have a 10GbE network switch that can be configured for jumbo frames06:11
twobitspriteand assuming all of your nodes have true 10GbE ports06:12
DrMaxI'm looking for prices and it varies wildly: from 200$ to ... 115k$06:12
DrMaxfrom 2000$ *06:13
DrMaxfrom 2000 to 115k$06:13
twobitspriteyeah, if you're doing this on your own dime, and/or limited budget, go with iSCSI over 10GbE if you can get the budget for that06:14
DrMaxNo, I have grant money06:14
DrMaxjust not an unlimited amount of06:14
twobitspritethen bilk them for all they're worth and spring for the FC rig! :P06:14
DrMaxSo, what does Fibre Channel does, beyond being stupid-fast?06:15
twobitspriteit's just stupid fast. that's it's entire purpose06:15
DrMaxwhat's the model? It creates shared memory spaces? Virtual FS? How does that work?06:15
twobitspriteit's for shared storage, as in disks. usually it links large RAID arrays06:16
DrMax(apparently you can route IP v4/v6 over it...?)06:16
DrMaxSo OS-side, I'd see what? Just a mount?06:17
twobitspriteyep06:17
DrMaxThat's interesting06:17
twobitspritea really fast mount shared by other systems. now, you have to use software to coordinate locks between other systems so you don't end up overwriting each other's blocks...06:18
twobitspritebut yeah, it just shows up as a block device06:18
DrMaxthat's be a cromulent, albeit expensive solution06:19
twobitspriteas in, you can `fdisk` the LUN, etc. Set up partitions, formate filesystems. just like a disk drive06:19
* twobitsprite googles "cromulent"06:20
twobitspriteahh06:20
DrMaxSo one plays host? or the shared volume can also be distributed?06:20
DrMaxcromulent is a made-up word from the Simpsons06:21
DrMaxit seems to mean "acceptable"06:21
twobitspriteDrMax: something has to be host, usually a "filer" like a NetApp or EMC appliance. It's basically a rack-mounted computer with tons of memory and high-speed networking card, running an OS that serves FC, iSCSI and/or NFS. There are opensource alternatives like TrueNAS, but you still need the hardware06:24
DrMaxis there a real advantage using the "real" appliances vs the hardware+opensource?06:27
DrMaxAlso, what kind of setup would you have to host a small-ish quantity of data?06:27
DrMax(small-ish is, say, 4 to 10TB, or something like that)06:28
twobitspritethe advantage is you have someone's desk to stand on when things go wrong. the enterprise stuff is about 10-50x the cost of the actual hardware. if you want to set your own filer by piecing together COTS hardware and configuring the OS yourself, go for it.06:29
twobitspriteas for the size of the data, the performance question is really about "spindles". in a RAID array, the number of actual drives is the most important part.06:31
DrMaxthat's the kind of thing I'd hire a grad student for06:31
DrMaxexplain?06:31
twobitspritesure, and that grad student can put it on their CV, so yeah06:31
DrMaxeverybody wins.06:32
twobitspriteDrMax: each drive as seek time (if it'f a physical disk, it has to wait for the spindle to come around and the reading head to position to the right cylinder, etc), so if you stripe the data across all the cylinders in the array, you can read from all of them at once instead of waiting for the next rotation06:33
twobitspritemodern SSD based system still have IO delays, but not as bad... but they're way more expensive.06:34
DrMaxit's usually in µs instead of ms06:34
twobitspriteoh, thanks for the correction, from the guy who doesn't know what FiberChannel is :P06:35
DrMaxI said I wasn't an admin, not that I never used a computer06:35
DrMaxon my current box, "seek" is 60µs06:36
twobitspritehe says from his ivory tower06:36
DrMaxthe view is nice from above06:36
twobitspriteDrMax: yeah, now too 20000 seeks on the filesystem at once06:36
twobitspritelet me know when you've run a production Orabel database for IBM06:39
twobitspriteOracle... it's late and I've been drinking. Good night XD :P06:39
DrMaxI also never said I'm masochistic06:39
DrMaxMaybe I should also run a km with glass shards in my shoes?06:40
twobitspriteBut you did come here asking for advice...06:40
DrMaxSoak my contacts in lemon juice?06:40
Mekaneckwhat has this to do with ubuntu support?06:40
twobitspriteMekaneck: nothing06:40
DrMaxMekaneck : does it have to?06:41
Mekaneckgo to offtopic if you want to chat06:41
Mekaneckit's a support channel after all06:41
twobitspriteMekaneck: go to bed if you want to dream about being in power. We're not interrupting on-topic chat, we're just chatting06:41
DrMaxthe conversation started about distributing tasks over linux boxes06:41
lotuspsychje!ot06:41
ubottu#ubuntu is the Ubuntu support channel, for all Ubuntu-related support questions. Please use #ubuntu-offtopic for other topics (though our !guidelines apply there too). Thanks!06:41
Mekanecktwobitsprite: in #ubuntu-offtopic you can just chat as you please, not here06:42
twobitspriteit's not like this is a busy channel and we're obstructing Ubuntu support questions06:42
Mekaneckdoesn't matter06:42
twobitspriteI'm sorry this conversation has bothered you06:43
Mekanecknot only me, others too06:43
lotuspsychjeyou can read the guidelines twobitsprite06:43
Mekaneckthey probably can't read06:43
twobitspriteyes, we're on IRC because we can't read06:43
lotuspsychjelets just move on please06:44
twobitspritethis conversation has been going on for at least an hour and no one has complained, and you wake up and want to exercise your domain as overlords06:45
lotuspsychjetwobitsprite: feel free to continue in #ubuntu-discuss06:45
twobitspriteyou started this discussion here06:45
Mekaneckand it ends here, enough said06:46
twobitspriteI think it's totally fair to have side conversations in a mostly dead channel in the middle of the night on Internet Relay CHAT as long as it's not interfering with on-topic discussions. If someone else joined and had a serious Ubuntu question, I would stop and address it06:47
lotuspsychjethe rules are the rules twobitsprite read the guidelines please06:47
lotuspsychjethe volunteers didnt invent the guidelines06:48
twobitspriteyes they did... did a corporate CEO invent the guidelines?06:48
lotuspsychjefeel free to dicsuss it in #ubuntu-discuss06:49
twobitspritenah06:49
DrMaxI'm off to bed too06:50
DrMaxtourlou06:50
twobitspriteg'night06:50
=== gabes5 is now known as gabes
=== pusher is now known as keypusher
BluesKaj'morning folks11:21
waltman2014:18
twobitsprite2114:19
lotuspsychjecan we help you waltman twobitsprite ?14:19
waltmanSorry, just a typo.14:23
semHi, I found a package in Debian that I can't find in Ubuntu; https://packages.debian.org/en/trixie/gnome-shell-extension-caffeine14:23
semcan't find it in ubuntu https://packages.ubuntu.com/search?keywords=gnome-shell-extension-caffeine&searchon=names&suite=resolute&section=all14:23
semis there a reason why it wouldn't be in the Ubuntu repos?14:23
semprevious search link was mistaken -- the one searching 'all' also finds no packages https://packages.ubuntu.com/search?suite=all&searchon=names&keywords=gnome-shell-extension-caffeine14:24
tomreynsem: it's available for some older ubuntu releases (only): https://paste.debian.net/plainh/e84c1b78 i don't know why, though. you can still install gnome-shell extensions through one of the gnome shell extension managers.14:36
semthanks -- it is a mystery!14:37
=== kelt0m is now known as fdp
DrMaxsem: because it was moved as a shell-extension: https://extensions.gnome.org/extension/517/caffeine/16:41
=== ancientz0 is now known as ancientz
=== pusher is now known as keypusher

Generated by irclog2html.py 2.7 by Marius Gedminas - find it at mg.pov.lt!