Network workload

Workload analysis

Figure 1 shows the relative throughput of the medium level network workload when executed as part of the combined workload set within a variety of differently constrained z/VM configurations.

Figure 2 shows the relative throughput of the high level network workload.

A general description of the elements shown in Figure 1 and Figure 2 is given in General description of relative throughput diagrams.

Observations for both levels of the network workload:

When processed during the reference execution, the relative throughput of the medium level network workload was about 30 % of that of the high level workload.

When less z/VM real memory was configured, the influence on the throughput of the network workload was mostly insignificant, except when 20 real CPUs were configured for z/VM, in which case the throughput slightly degraded as less real memory was configured for z/VM.

With 15 or more real CPUs, the throughput of the medium level network workload was consistently higher than that achieved during the reference execution, except when 64 GiB of real memory was configured for z/VM.

When 15 real CPUs were configured for z/VM, the high level network workload regained some throughput with 96 GiB real memory or less.

The throughput of the network workload was differently affected by the number of real CPUs configured for z/VM, depending on the workload level:

• The medium level network workload achieved better than reference throughput with 15 or more z/VM real CPUs configured. The throughput significantly declined as 10 or less z/VM real CPUs were configured.
• Within any configuration, the high level network workload did not reach the throughput value achieved during the reference execution. The throughput break already occurred with 15 z/VM real CPUs configured, down to about 40 % of the reference throughput, and then about declining linearly with decreasing numbers of configured real CPUs.

Conclusions for both levels of the network workload:

The network workload is primarily impacted by z/VM real CPU constraints.

The virtual systems (client and server) configured with two virtual CPUs processing the high level network workload is capable to achieve three times the throughput of the virtual systems configured with one virtual CPU processing the medium level network workload. However, as soon as noticeable constraints are introduced at the z/VM level, throughput of the high level network workload degrades significantly. Opposed to that, as long as z/VM resource constraints are not too excessive, the virtual system that is processing the medium level network workload is able to retain and partially even exceed its reference throughput.

CPU usage analysis

Figure 3 shows the average CPU load for the medium level network workload when processed as part of the combined workload set within a variety of differently constrained z/VM configurations.

Figure 4 shows the average CPU load generated by the high level network workload.

A general description of the elements shown in Figure 3 and Figure 4 is given in General description of page distribution and page read rate diagrams, except that in Figure 3 and Figure 4 CPU load is presented instead of throughput.

Observations for both levels of the network workload:

The CPU load generated by the medium level network workload was mostly independent from the amount of z/VM real memory.

The CPU load generated by the high level network workload was mostly independent from the amount of z/VM real memory when 10 or less real CPUs were configured for z/VM. When 20 z/VM real CPUs were configured, the CPU load gradually declined with the amount of z/VM real memory, and when 15 real CPUs were configured, the CPU load gradually increased with the amount of z/VM real memory.

The CPU load generated by the server systems generally was a little bit higher than that generated by client systems, regardless of the workload level.

Conclusions for both levels of the network workload:

The CPU load analysis confirms the diagnosis of the workload analysis, that the throughput of medium level network workload is affected later than that of the high level network workload when real CPU constraints are introduced at the z/VM level. It seems that in z/VM real CPU constrained situations the high level network workload gained advantage with respect to CPU costs by being able to use a second virtual CPU.

Memory usage and paging workload analysis for CID 64/20

As usual, the detailed memory usage and workload analysis is only presented for the z/VM configuration with 64 GiB real memory and 20 real CPUs (CID 64/20).

Figure 5 shows the average page distribution and the average page read rate for the medium level network workload when processed as part of the combined workload set when z/VM is running in a CID 64/20 configuration.

Figure 6 shows the average page distribution and the average page read rate for the high level network workload.

A general description of the elements shown in Figure 5 and Figure 6 is given in General description of page distribution and page read rate diagrams, except that in Figure 5 and Figure 6 CPU load is presented instead of throughput.

Observations for both levels of the network workload:

Both levels of the network workload exhibited a rather small amount of instantiated memory that was backed by resident memory only to a small fraction (about 13-14 %).

The z/VM paging activity caused by the network workload was negligible.

Conclusions for both levels of the network workload:

The network workload is not a big memory consumer. Nevertheless, the high level network workload was affected by z/VM memory constraints when z/VM was configured with 15 or 20 real CPUs.

A possible explanation for the low memory consumption of the network workload might be that the workload only makes use of small data buffers. These buffers are initially allocated once, and then repeatedly used for sending and receiving data, that is, for transferring data to/from the Linux device drivers. In that respect the network workload behaved similar to the direct I/O variant of the file system I/O workload.

Once received, data is discarded and not copied for further use, as it would be the case in a real application. Insofar, the results obtained for the network workload should be considered as a metric for the overhead that results from transferring data between virtual systems, but not as a metric that also covers the further processing of the transferred data.

With that in mind, the high level network workload might be a good candidate for reserving memory by means of the z/VM set reserved command, in order to support z/VM in keeping most of the small but highly used data transfer buffers in z/VM real memory.