Standard TCP / UDP Tests

The goal of these tests are to check the general performance of the hosts network combination by running some standard TCP and UDP tests. Flows from both types are generated between two sets of DAS-2 cluster nodes which are separated in two VLAN's. To obtain sufficient diversity in the # flows, each node in the source VLAN is sending a flows to each of the N receiving nodes in the destination VLAN. The flows are created with the Iperf traffic generator that is also capable of running multiple parallel flows implemented with the pthread library.

As explained in the "Description" section, TCP tests were generated between both VLAN's. Because the RTT is very small large variations of the TCP window make no sense here. The tests were performed from 1 to 12 source - destination host pairs, where each source host is sending parallel a flow to all receiving hosts. The used TCP window per flow will be divided by the # parallel flows. The TCP window sum, taken over all parallel flows, is varied from 32 KByte to 256 KByte and the number parallel flows is varied from 1 to 12. As output Iperf interval bandwidth reports of the individual flows are stored with an interval time of 4 s.

As a result the TCP throughput sum, taken over all source - destination flows, is analysed which is calculated by taken the median values of the bandwidth reports that are re-sampled for this purpose in exactly 4 s. intervals. In the sum throughput values are displayed as a function of the TCP window sum that has been taken here over all flows.

From there follows that in almost all situations the TCP throughput is above about 980 Mbits/s. Only for low TCP window values in combination with a few # flows the TCP throughput is lower.

UDP tests has been performed as a function of the shaping bandwidth (variation: 50 Mbits/s, ..., 950 Mbits/s), the # node pairs (1, ..., 4 node pairs), and the # parallel flows per node pair (1, ..., 5 parallel flows). Concerning the from these tests obtained bandwidth and packets lost values it appeared that the loop-back capacity could be well utilised. However, especially with multiple (parallel) flows, dependent from the used shaping bandwidth, some 10 - 20% of the connections got broken. If we have to deal here with network and / or host effects is still subject of further investigation.