CLUSTER OPERATIONS · SCHEDULERS · SCIENTIFIC COMPUTING
Auto-healing is a mechanism, not an objective. At 10,000 GPUs failure is the steady state, and every healing decision trades detection latency against the cost of acting wrongly — including on the node that was fine.
The backdrop of my homepage is a real Hodgkin-Huxley connectome with spike-timing dependent plasticity, and the scheduler widget is a working backfill loop. Here is how both run without costing anything at first paint.