Setup

• All machines were dual-core Xeon 3060 machines
• All machines were hooked up to a single gigabit switch
• ZooKeeper 3.3.0 was used for both the clients and servers
• 3 machines made up the ZooKeeper cluster
• The benchmark utility is a Java client that will connect to a specific ZooKeeper server (always the same) and do n back-to-back reads or writes of a single object of a given length

Latency with benchmark client running on the ZooKeeper server

Observations during reads:

• ZooKeeper takes about 20-25% of one core
• The benchmark client uses about 10% of the other
• The kernel uses about 20-25% of each core
• The userspace programs spend most of their time waiting on a futex.
• The replicas are idle.

Observations during writes

• ZooKeeper takes about 50% of the CPU
• The kernel uses about 7% of the CPU
• The first replica uses about 60-80% of its CPU on ZooKeeper and about 12% on the kernel
• The second replica uses about 40-60% of its CPU on ZooKeeper and about 12% on the kernel

Latency with benchmark client running on a separate machine

Throughput

The maximum read throughput I was able to observe was ~50,000 512Byte-read operations per second. This was using 8 processes (without my own Java threads) on each of 4 client machines. This does not entirely saturate the cores on the server – they are each between 5% and 30% idle at any given time.

The maximum read throughput I was able to observe from a single client machine was only ~29 512Byte-read operations per second using 8 processes (each of 4 Java threads). I don't understand the behavior of the client library (which seems to spawn about 11 threads of its own), so I can't really explain this.

1 client machine: 5.6Kops/s

2 client machines: 10.4Kops/s

3 client machies: 15.7Kops/s

4 client machines: 20.0Kops/s