On 18/09/25 at 21:18 +0200, Lucas Nussbaum wrote: > On 18/09/25 at 16:54 +0200, Santiago Vila wrote: > > On Thu, Sep 18, 2025 at 03:51:09PM +0200, Lucas Nussbaum wrote: > > > > > > Is the build failure you got on arm64 similar to that, or it's really > > > > different? Can you quote the part of the build log where available > > > > memory and swap are shown? > > > > > > There's 32 GB of RAM on the nodes I was using, so this should not be a > > > problem. > > > > We don't really know for sure. > > > > If you look at the debian/rules file: > > > > https://salsa.debian.org/python-team/packages/graph-tool/-/blob/master/debian/rules?ref_type=heads > > > > you will see that there is some code to calculate the number of CPUs > > to be used based on available memory. It could still be the case that > > we need to adjust that for gcc-15. > > > > For example, if the appropriate value for MEM_PER_CPU when using > > gcc-15 happened to be larger than 16000, then building with 2 CPUs > > might not work if you have just 32 GB of available memory. > > > > Can you try with *some* swap? When using AWS instances, you will need a > > swap file, so you can do something like this: > > > > fallocate -l 8G /boot/swap > > chmod 600 /boot/swap > > mkswap /boot/swap > > echo "/boot/swap none swap sw 0 0" >> /etc/fstab > > swapon -a > > > > [ Also: Make sure you are not trying to build several packages at once, > > because in such case the calculation above would be quite useless ]. > > Hi, > > I did not look deeper yet, but I could reproduce it on amd64 (while > building libgraph_tool_generation_la-graph_merge.lo)
I retried on a machine with very few parallel builds and an huge amount of RAM (6 TB), and it also failed.
