SMC-R Protocol Strengthens Data Sharing Performance While Reducing CPU Costs

In z/OS, a significant emphasis was placed on usability. RoCE features will not require additional software configuration or individual operational procedures. Instead, they “inherit” the OSA-Express definitions, such as VLANs, and mimic OSA operations, including starting and stopping OSA interfaces.

Performance Gains

So what performance gains are possible with this protocol? One is significant latency reduction for workloads with a chatty request/response message patterns, such as application servers to database servers. It also offers a significant reduction in CPU cycles consumed by the networking stack for workloads with streaming data patterns where large amounts of data flow in one direction, as with FTPs.

To measure the performance benefits of SMC-R, IBM completed several micro-benchmarks using the IBM Application Workload Modeler (AWM). Micro-benchmarks focus on exercising the entire networking path from a sockets-application perspective (TCP socket to TCP socket), but with no application logic present to process the data and perform business-related processing. As a result, the benchmarks accurately depict the best possible performance benefits for the SMC-R protocol. With real workloads, these benefits will likely be smaller once application-specific processing and associated latency are introduced. Micro-benchmarks comparing SMC-R using the 10GbE RoCE Express feature to TCP/IP using 10GbE OSA Express5 indicated results upwards of:

  • An 80 percent drop in network latency for request/response traffic patterns
  • A 60 percent decline in networking stack-related CPU cycles
  • A 60 percent increase in network throughput for streaming workloads

These benefits were evident on both server and client hosts, demonstrating the efficiency of SMC-R and RoCE in moving data across the network. The sender can move data more efficiently while the receiver can receive data in a big “chunk,” getting interrupted less often to receive packets. The receiver also generates significantly fewer acknowledgements for the sender to process. Both client and server sides of these benchmarks were performed on a zEC12 with dedicated OSA-Express and RoCE Express features per LPAR connected via a single 10GbE Ethernet switch.

Ultimately, the combination of SMC-R and the 10GbE RoCE Express feature delivers compelling improvements in network performance with reduced latency, improved throughput and CPU cost while providing application transparency.

Gus Kassimis is a System z networking architect and designer in z/OS CommServer and one of the lead architects of the SMC-R protocol.

Jerry Stevens is a senior technical staff member with IBM Software Group and works in AIM Enterprise Networking Solutions Architecture Strategy and Design with a focus on communications hardware architecture. He has more than 25 years of experience with z/OS network communications.

comments powered by Disqus



2019 Solutions Edition

A Comprehensive Online Buyer's Guide to Solutions, Services and Education.

Expanding Beyond Borders

IBM zEnterprise System announcement delivers a new dimension in computing.

Reducing Complexity

The next-generation System z server is more than just a fast, scalable solution

Address the IT Conundrum

A Smarter Computing-based infrastructure enables a smarter planet

IBM Systems Magazine Subscribe Box Read Now Link Subscribe Now Link iPad App Google Play Store
Mainframe News Sign Up Today! Past News Letters