Summary of Key Findings
This section summarizes the performance evaluation of z/VM 3.1.0, including TCP/IP Feature for z/VM, Level 3A0. Measurements were obtained for the CMS-intensive, VSE guest, Telnet, and FTP environments on zSeries 900 and other processors. For further information on any given topic, refer to the page indicated in parentheses.
z/VM 3.1.0 includes a number of performance enhancements (see Performance Improvements). Some changes have the potential to adversely affect performance (see Performance Considerations). Lastly, a number of changes were made that affect VM performance management (see Performance Management):
- Performance Improvements
- 64-Bit Support
- VCTC Pathlength Reduction
- Miscellaneous CMS Improvements
- Gigabit Ethernet Support via QDIO
- Performance Considerations
- MDC Tuning with Large Real Memory
- Large V=R Area
- Performance Management Changes
- Monitor Enhancements
- CP Control Block Changes
- QUERY FRAMES Command
- NETSTAT Command
- Effects on Accounting Data
- VM Performance Products
Migration from VM/ESA 2.4.0 and TCP/IP 320:
Benchmark measurements show the following performance results for z/VM 3.1.0 (31-bit CP and 64-bit CP) relative to VM/ESA 2.4.0:
CMS-intensive | Throughputs and response times were equivalent for all 3 cases. Processor requirements for VM/ESA 2.4.0 and z/VM 3.1.0 with 31-bit CP were equivalent. Processor requirements for z/VM 3.1.0 with 64-bit CP increased by 0.8% to 1.8% for the measured environments as a result of the 64-bit support (see CMS-Intensive). |
VSE guest | The performance of all 3 cases was equivalent for both the V=R and V=V environments (see VSE/ESA Guest). |
TCP/IP stack machine processor requirements for TCP/IP 3A0 decreased by approximately 1% relative to TCP/IP 320 for the measured Telnet and FTP workloads (see TCP/IP).
CMS measurements using the z/VM 3.1.0 64-bit CP running on a 2064-1C8 processor with 8G total storage showed an internal throughput rate (ITR) improvement of 4.6% when most of that storage was configured as real storage as compared to 2G real storage and 6G expanded storage. For 12G total storage, the ITR improvement was 3.8%. Additional measurements in these storage configurations show that it is important to reassess minidisk cache tuning when running in large real storage sizes. Finally, measurements are provided that help to quantify the amount of real storage that needs to be available below the 2G line when VM is run in large real storage sizes. While CP now supports all processor storage being configured as real storage, it is still recommended that some storage be configured as expanded storage (see CP 64-Bit Support).
Measurement results on a 9672-ZZ7 processor demonstrate the ability of the new QDIO support to deliver Gigabit Ethernet throughputs of up to 34 megabytes/second using a 1500 byte packet size and up to 48 megabytes/second using an 8992 byte packet size. The primary limiting factor is TCP/IP stack machine processing requirements, suggesting that even higher throughputs are achievable on a zSeries 900 processor or if 2 or more stack virtual machines are used (see Queued Direct I/O Support).
The data privacy provided by the Secure Socket Layer support increases processor requirements relative to connections that do not use SSL. For connect/disconnect processing, 10x to 28x increases were observed for new sessions, while 6x to 7x increases were observed for resumed sessions. Increases ranging from 4x to 10x were observed for an FTP bulk data transfer workload, depending on the cipher suite used (see Secure Socket Layer Support).
The Linux IUCV driver sustains significantly higher data rates relative to using virtual channel-to-channel (VCTC) through the Linux CTC driver. Measurement results show 1.4-fold to 2.4-fold increases, depending upon data transfer size. These higher throughputs are due to lower processor requirements (see Linux Guest IUCV Driver).
VCTC performance has been significantly improved by VM/ESA 2.4.0 APAR VM62480, now part of z/VM 3.1.0. With this improvement, VCTC processor requirements are similar to real ESCON CTC. Throughput for bulk data transfer is 2.4 times higher than real ESCON CTC for the measured environment due to the absence of real CTC latencies (see Virtual Channel-to-Channel Performance).
Measurement results demonstrate the ability of TCP/IP Telnet to support 5100 CMS users with good performance. Relative to the corresponding VTAM support, however, response times and processor usage were higher due to increased master processor requirements (see Migration from VTAM to Telnet).