Re: Ampere EMAG
On 11/15/20 2:30 PM, Arnd Bergmann wrote:
> On Sun, Nov 15, 2020 at 12:36 PM Christian Kastner <email@example.com> wrote:
> The SolidRun HoneyComb LX2K is now advertised as supporting
> m.2 and PCIe, plus up to 64GB of RAM on a 16-core SoC, which is a
> nice ratio between memory and CPU.
Wow, this one is even (relatively) cheap. I'll definitely have to take a
closer look at this, thanks!
The production description says most of kernel, u-boot upstreaming is
still work in progress, but I found a page that supposedly got a recent
Ubuntu running on it, so this looks pretty good already.
>> I recently rebuilt about ~1000 packages and overall parallelization was
>> less than I hoped for (but not really surprising: after all, the package
>> build sequence is serial, and the build+test stage, where
>> parallelization would shine, often quite short).
> I suppose for large-scale rebuilds it ends up being ideal to have
> each CPU build one package in a container and not rely on
> on spreading out one package across multiple CPUs, but this
> needs even more RAM if you want to do the builds in virtual
> machines instead of containers.
That's how I do it on amd64 currently. I assign each worker VM 4 cores
and 8GB of RAM, and let a few workers run in parallel. I'm sure 8GB is
insufficient for some packages (eg: the kernel), but I haven't run into
one of them yet for my specific needs. And it's just a changeable parameter.
My mid-term goal is to get most of the package builds done in
containers, but I have yet to analyze the toolchains for that. I'm
currently favoring sbuild+autopkgtest+lxc (simply because I'd like to
stick to sbuild+autopkgtest), but there are at least a half-dozen
alternative implementations.  has some, but the list is incomplete.
In any case, I can't get rid of the VMs entirely, as some packages
operate closer to the system. For example, maintaining the keyutils
package, I ran into platform-specific bugs in the kernel, which I can't
test on porter boxes.