[HBASE-3115] HBaseClient wastes 1 TCP packet per RPC

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 0.20.1, 0.20.2, 0.20.3, 0.20.4, 0.20.5, 0.20.6, 0.89.20100621
• Fix Version/s: 0.90.0
• Component/s: Client
• Hadoop Flags: Reviewed
• Tags: performance
• Assignee: Todd Lipcon
• Reporter: Benoit Sigoure
• Created: 15/Oct/10 08:25
• Updated: 20/Nov/15 12:42
• Resolved: 19/Oct/10 06:04

Description

In ipc/HBaseClient.java, the method sendParam does:

out.writeInt(dataLength);      //first put the data length
out.write(data, 0, dataLength);//write the data

While analyzing some tcpdump traces tonight, I saw that this consistently translates to 1 TCP packet with a 4 byte payload followed by another TCP packet with the RPC itself. This makes inefficient use of network resources and adversely affects TCP throughput. I believe each of those lines translates to a write system call on the socket's file descriptor (unnecessary system calls are also bad for performance). The code attempts to call out.flush(); but this approach is ineffective on sockets in Java (as far as I empirically noticed over the past few months).