It would be better if you used ScyllaDB Manager for the job as it knows to handle repair parallelism and intensity with ease, but I see you have a 20-node cluster.
That said, you can repair multiple nodes at the same time. You could also simply run nodetool repair once and let it eventually complete. However, you’ve already said you currently are having write failures, so ideally you’d rather check on these before adding more background work to your running cluster.