Rich Freeman <rich0 <at> gentoo.org> writes:
> Don't get me wrong - I appreciate the desire for bare-metal > performance in the high-performance computing world. I've heard > stories/rumors of Gentoo getting attention elsewhere in this domain, > and we have a disproportionate number of physical scientists and such > in the community (including probably half of the Council - we joke > about it). I've even heard of Gentoo used in high-throughput trading, > though a lot of that has moved on to ASICs and such and nobody talks > openly about what they're doing. Yep; lots of folks are putting their *nix expertise into FPGAs these days as a way to protect "their Intellectual Property". Here's a prime example in the drug discovery world [1]. Trading with Gentoo:: Yep. I was hustle via a NYC head_hunter for several projects, some years back, but they would never disclose the companies. One was some wealthy individual. They wanted to emcumber me before they told me anything; not a good sign, besides I'm not too fond of NYC. The more bad stuff I told the HH about myself, the more they liked me as candidate. The pay scale was way to high for my abilities anyway... Yep:: nobody talks..... lots of real wise guys in NYC. > I was just trying to point out that containers are very different from > VMs, while generally trying to solve the same sorts of problems. VMs > create continuous execution overhead and are memory-expensive. > Containers have zero execution overhead and are very memory-efficient. > Of course, if you throw 5x as many running processes on the same PC > you're still going to consume more RAM and CPU, but 5 containers > running on 1 PC tend to be pretty close to the CPU+RAM requirements of > linux hosts running on 5 PCs. If you're just using containers for > configuration-management/etc and just run one container on a node, > then you're going to be very close to the same performance you'd get > running it on bare metal. VM are obsolete compared to containers, when you start looking closely at timing and latencies which then effects throughput. That's pretty much accepted mathematically by virtually all of the clustering devs I interact with. It does not mean VMs are dead or not useful, but they are not in the competition any more on performance driven needs. Look Rich. Believe me, when you say things I listen. It's on my to do list to evaluate CoreOS vs bare metal. Not to beat a dead horse but I do need a fully unattended install semantic to do the regression testing for routine cluster needs and my "half baked" ideas.... I do not believe regression test results in vm or container setups. Maybe the first or second digit of accuracy. I'm old school and I have to isolate things on hardware. That's just how I roll:: I guess it's the EE in me. Trust but verify...... So yes at some point I intend to vet the CoreOS thing, as it is very close to gentoo with ebuilds and such.... I think I'm the one that pointed coreos out on the gentoo user list; some time ago, as a derivative or rip-off of gentoo..... Folks said ChromeOS was from Gentoo and CoreOS was from ChromeOS.... (ring any bells?). > From the kernel's perspective every linux system uses containers. > They just tend to use a single container. The kernel doesn't do > anything differently when a process spawns in a container. When that > process looks out at the world the kernel shows it everything within > its namespaces. That is true whether you have one set of namespaces > on the system or 50. As far as I'm aware the system calls all take > just as long to run either way. Containers really are just about > adding one more field to the keys in various kernel objects like > processes/tasks. WE have kernel shark (via trace-cmd) now and heaptrack too. Those (2) tools alone should let you gather actual data on what you have stated above and publish it. If you want a bunch of links to kernel shark info and examples just let me know. What would be keen (and is on my todo list) is to take kernelshark and use it for some deep analysis work on gentoo. Then publish a gentoo wiki page on KernelShark so the community can see a cool example on Gentoo. Kernelshark bridges that kernel/OS barrier and can quantify the actual timing and latencies and problems in a full stack or particular layer of the stack. It is addicting and can consume days of your time, just so you know (in advance). On another note:: What I'm missing (and it's definitely new learning material to me) is a robust, flexible DAG tool(s). What do you know about for DAG and such tools/codes? James [1] https://www.deshawresearch.com/