GitLab

A bit too short for a fast qemu timing.

this allows the shared memory regions to be created in /dev/shm/ e.g.
instead of the work dir.

This breaks large mtu experiments, but when proxied makes a huge
difference. Should make this a runtime parameter.

Can now select between sockets (default) or rdma.

Instead of static ordering of simulator types. This simpliefies the
experiment runners but also simplifies adding additional features.
Immeadiate goal is adding network-to-network connections

These were too large until now.

Otherwise on a full queue we might overwrite entries repeatedly causing
all sorts of mayhem.

It's accessed from both poll thread and IO thread

Because dev/net_info are packed gcc does not seem to align them
properly.

Only show one message when we start to wait, and one when we succeeded.

Just memcpy risks the received reading incomplete messages.

This allows NetListen to be done before other intialization, so that
listening sockets for peers are created already. This makes
orchestration a lot easier. E.g. listener proxy will have listening
sockets for peers already created when still waiting for connector proxy
to connect.

Preparation to add additional proxy implementations.

Uses a TCP socket to exchange connection info.

To add manual infiniband initalization as an alternative in a next step.