Generation of workloads to test RAMCloud

This document defines a simple plan to evaluate the performance and scalability of RAMCloud. The first section lists the dimensions we need to take into consideration and the final measurements we should end up with. The third section has ideas for simulated workload situations under which these measurements could be made in the future.

Measurements of interest

The ordering indicates relative importance in the lists that follow. Only the first 2-3 are likely to be attempted to start things off.
At first these measurements could be made on a simple set-up with a single node running master/coordinator/backup/client.

Client's perspective

1. Latency under no-load
   Input parameters - Object Size, Type of operation (Object vs Table, Read vs Write), Capacity of RAMCloud in use, Master failures, Object access distribution (Zipf?)
   Output measured in - micro seconds
   Measured end-to-end from the perspective of the client.

2. Throughput of operations
   Input parameters - Object Size distribution, Type of operation, Object access distribution, Capacity of RAMCloud in use, Master failures
   Output measured in - number of object reads per second, number of object writes per second

3. Throughput of IO
   Input parameters - Object Size distribution, Type of operation, Object access distribution, Capacity of RAMCloud in use, Master failures
   Output measured in - number of client bytes reads per second, number of client byte writes per second (ideally should only include real content size and not count the overhead imposed by RAMCloud structures)

4. Maximum simultaneous Connections
   Input parameters - Number of masters, Number of client processes, Number of client machines
   Output measured in - number of connections

5. Latency under load
   Input parameters - Number of operations under load, Object Size, Type of operation (Object vs Table, Read vs Write), Capacity of RAMCloud in use, Master failures, Object access distribution (Zipf?)
   Output measured in - micro seconds
   A pre-defined workload is first imposed on the system in terms of number of simultaneous operations. Latency measurements are then made for each type of operation.

Of internal interest

1. Number of rpcs

2. Internal Hash-table size

3. Actual DRAM used versus raw DRAM required for object contents

Graphs of interest

Round 6 - InfRC - (generated with code version - commit 24f6d2)

• New cluster with 36 nodes total used.

• master/coord at rc01.rccluster. clients on all other nodes - rc01-36.rccluster

• latency and throughput curves measured for different objects sizes read under load (both quantities are measured in the same experiment)

• only reads are measured - single object within a single table is read

• Object sizes at 100, 1000 and 10000 bytes

• transport=infrc, Throughput versus load
  

  Open

  

• transport=infrc, Latency versus load
  

  Open

  

Round 5 - InfRC - (generated with code version - commit 24f6d2)

• New cluster with 33 nodes total used.

• master/coord at rc01.rccluster. clients on all other nodes - rc01-36.rccluster

• different objects sizes

• load vs time-per-operation measurements only.

• transport=infrc, Operations per second (throughput) versus load with 1000 byte objects
  

  Open

  

• transport=infrc, Operations per second (throughput) versus load with 100 byte objects
  

  Open

  

Round 4 - InfRC - (generated with code version - commit db388ad3)

• Reference measurement - Steve's access of 100 byte objects on single server - 7000 ns

• Single server round trip time - 100,000 measurements - 100 byte objects in 4833 ns run-bench script

• No load - 100,000 measurements - master/co-ord on rc01, client on rc02 - 100 byte objects in 5109 ns run-perf script

• No load - 10,000,000 measurements - master/co-ord on rc01, client on rc02 - 1000 byte objects in 5093 ns run-perf script

• 10.0.0.x NICs were used.

• 1 hour passes - cluster is as idle as the last experiment. But.....

• Single server round trip time - 100,000 measurements - 100 byte objects in 9869 ns run-bench script

• No load - 100,000 measurements - master/co-ord on rc01, client on rc02 - 100 byte objects in 9878 ns run-perf script

• transport=infrc, Operations per second (throughput) versus object size
  

  Open

  

Round 3 - (generated with code version - commit db388ad3)

• Co-ordinator and Server are both running on rc01

• Client making timing measurements is at rc03 - the queen

• Load - additional clients are run on rc04-6. If load > 3 is required, multiple clients are run on the same node. The Workers.

• Builds are with DEBUG=no unless otherwise noted.

• Object size is fixed at 1000 bytes.