Existing Practice

• Compute clients with pending input
• Read 4K of requests from each ready client
• Loop: Execute 10 from each ready client until all are empty
• Wait for more input