Evaluation clusters
Local cluster
The image above shows the physical topology of our small dedicated test cluster. The cluster comprises of seven machines, connected via a 1G switch. Detailed information about machine hardware is in the table below.
Host hardware
| Host | Name | Architecture | Cores | Threads | Speed (GHz) | Memory (GBs) |
|---|---|---|---|---|---|---|
| H1 | freestyle | Intel Gainestown | 4 | 8 | 2.26 | 12 |
| H2 | hammerthrow | Intel Gainestown | 4 | 8 | 2.26 | 12 |
| H3 | backstroke | Intel Gainestown | 4 | 8 | 2.26 | 12 |
| H4 | tigger | AMD Magny Cours | 48 | 48 | 1.9 | 64 |
| H5 | uriel | AMD Valencia | 12 | 12 | 3.1 | 64 |
| H6 | michael | Intel Sandy Bridge | 12 | 24 | 1.9 | 64 |
| H7 | raphael | Intel Sandy Bridge | 12 | 24 | 1.9 | 64 |
EC2 cluster
For many of our experiments, we used large EC2 clusters of 100
m1.xlarge instances. We also used smaller sub-clusters of
16, 32 and 64 instances for back-end execution engines that only support
power-of-two node counts (e.g. PowerGraph).
All EC2 clusters ran in the us-east-1a availability zone.
We mostly ran our experiments during US night time in order to reduce
the variability in results (which can be up to +/- 100% in makespan at
other times).
VM specification
| Count | Type | vCPUs | Memory (GBs) |
|---|---|---|---|
| 100 | m1.xlarge | 4 | 15 |