Virtual Channel-to-Channel Performance
This section covers two separate aspects of VCTC performance. First, results are presented that quantify a recent VCTC performance improvement. Second, real ESCON CTC is compared to VCTC with this improvement included.
The same basic methodology was used for both evaluations. A VM/ESA 2.4.0 system was configured with two VCTC-connected V=V guest virtual machines. The VM system was run in an LPAR with 5 shared processors on a 9672-R55. The workload consisted of a binary FTP get of a 10M file from one guest to the other. The transfer was memory-to-memory (did not involve any DASD I/O) because the source file was resident in the minidisk cache and the target file was defined as /dev/null. CP QUERY TIME output was collected for both guests before and after the FTP file transfer. Three trials were obtained for each measured case. The results were quite repeatable; representative trials are shown in the results tables.
VCTC Performance Improvement
This improvement significantly reduces the amount of processing
required by CP's virtual CTC implementation by improving the efficiency
with which the data are copied from source to target virtual
machine. This improvement was first made available in VM/ESA 2.4.0
through APAR VM62480 and has now been incorporated into z/VM 3.1.0.
Measurements made without and with APAR VM62480 are summarized
in Table 1.
Table 1. VCTC Performance Improvement
VM62480 runid | no GLLVCTC7 | yes GLLVCTC6 | ratio
|
---|---|---|---|
elapsed time (sec) Megabytes/sec CP CPU time virtual CPU time total CPU time | 0.850 11.76 0.59 0.37 0.96 | 0.517 19.34 0.28 0.36 0.64 | 0.61 1.64 0.47 0.97 0.67 |
Note: binary FTP get of a 10M file to /dev/null; client and server on separate V=V guests connected by VCTC; MTU=32760; VM/ESA 2.4.0; 9672-R55; LPAR; 5 shared processors |
The results show a 53% reduction in CP CPU time and a 64% increase in throughput.
Comparison of Real ESCON CTC to Virtual CTC
The VCTC connection between the 2 guests was then replaced
with a real ESCON CTC connection and the measurement was repeated.
Table 2 compares these results to the VCTC
results for the improved case shown in the previous table.
Table 2. Comparison of Real ESCON CTC to Virtual CTC
CTC type runid | real GLLRCTC2 | virtual GLLVCTC6 | ratio
|
---|---|---|---|
elapsed time (sec) Megabytes/sec CP CPU time virtual CPU time total CPU time | 1.220 8.20 0.28 0.36 0.64 | 0.517 19.34 0.28 0.36 0.64 | 0.42 2.36 1.00 1.00 1.00 |
Note: binary FTP get of a 10M file to /dev/null; client and server on separate V=V guests connected by real ESCON CTC or VCTC; MTU=32760; VM/ESA 2.4.0; 9672-R55; LPAR; 5 shared processors |
The results show that both cases have equivalent processing efficiency but the VCTC case achieved 2.36 times higher throughput because the latencies associated with real ESCON CTC are eliminated.
To be eligible for IOASSIST, the VM system would need to be run on a basic mode processor (not in an LPAR) and the two guests would need to be run in V=R or V=F virtual machines. If IOASSIST had been in effect, the virtual CTC results should be essentially the same but the real CTC results should differ in the following ways:
- elapsed time: little change
- MB/sec: little change
- CP CPU time: decrease
- virtual CPU time: little change
- total CPU time: decrease
We would expect little change to elapsed time and MB/sec because these are primarily determined by the real ESCON CTC latencies rather than by the processor usage.