How Does S390x Performance Compare To X86_64?

Mostly I’m a wonk trying to make sense of it all. s390x, the IBM Z 64-bit architecture, is a big machine designed for a large number of highly predictable transactions with high reliability and a number of hardware services enabled. In other words, it’s ideal for huge amounts of work that are mostly short transactions with occasional very large ones (e.g., Database OLTP). The operating system (e.g. z/OS), the virtualization layer (e.g. z/VM), and most applications are written to run on this type of architecture. In the benchmarks you read about, s390x tends to get the highest number of transactions per chassis and thus is optimized for running special services (e.g. zIIPs for debuggers) and I/O services provided by the mainframe itself. Also, the number of transactions and services to keep the machine up 24×7/365 is largely beyond human capacity. Thus a different meaning of “performance” is used than pure speed of transactions.

In contrast, x86_64 CPUs from Intel and AMD are basically the default workhorses of the computing world. They have a high clock speed and can give a single thread a big speed boost, and are supported by a massive amount of software and by an enormous developer base and hundreds of millions of computers, all based on the instruction set. So for single-threaded work, for developing software on a computer, for responding quickly to user input, for running most open-source software out of the box, x86_64 is fast. And it is easy to optimize for. The standard set of compilers, system libraries, and pre-optimized packages for x86_64 are all very mature and frequently updated to take advantage of the latest processor models.

When actually tuning systems to perform better, one typically can either opt for high, constant performance 24/7 (good for OLTP servers and their big transactional workloads) and hardware crypto accelerators, plus high density virtualization (which makes z/VM very powerful for IBM’s mainframe server) – or go for very fast scaling and a huge software repertoire at low cost for i.e. small “microservices” (a la “cloud native” software), running in many small instances that collectively serve lots of requests – so x86_64 CPU’s in cheap hardware (even consumer hardware) at low cost /core are just a lot better than high-end mainframe CPU’s at such workloads, in my experience. Endianness of s390X has to be watched too as a special case when porting. Specialty hardware for e.g. very fast crypto can give surprising efficiency even per transaction in the end for high-transaction workloads – as long as peak performance per thread is not needed.

I have one JVM app running on s390x and the Crypto code runs much faster there since it’s all done in hardware on the chip. For my Microservice farm, running on x86_64 is way cheaper upfront, and it scales much better that way. To get a feel for the two architectures I’d run your critical path through a test setup and then look at latency distributions, throughput under load, and operational overhead of both to get a real feel for what’s going on.

This answer is short and to the point for your reference. There are different workloads and use cases that are better suited for s390x and x86_64. s390x performs best with constant high volume enterprise workloads (huge volumes of transactions) and has built-in crypto and data compression as well as RAS features that help to prevent downtime. s390x also scales vertically extremely well with very high virtualization density and special processors that can offload database and crypto work at scale resulting in higher efficiency per transaction. x86_64, on the other hand, tends to perform better with single-threaded workloads, as well as better for developers and cost-sensitive cloud-native applications due to higher clock rates, better compiler and optimizer support as well as a much larger set of development tools.

Portability / dev issues: s390x is nasty to run binaries on (due to endianness and fewer pre-compiled packages for most distributions), but decent Linux support for the architecture is available from the big distributions. Measuring P95/P99 latency (and associated throughput) plus operational complexity is typically more meaningful than headline numbers for your particular use case. I personally run s390x for high-volume, mission-critical transaction servers and x86_64 for highly flexible, containerized or cost-sensitive workloads. I prototype both to see which is actually faster for me in the end.

Mainframe vendors say “nobody ever got fired for buying IBM” but for them performance is a different kind of performance metric - career performance. The perceived risk of migrating a core banking system from s390x to x8664 is so high that one would have to experience performance improvements of orders of magnitude to even consider such a move. The mass of the installed base, the risk-adverse nature of industries that run such systems and the huge complexity of such moves all create a sort of force field around such systems. In such cases “performance” is does it perform well enough not to incur a huge and potentially career-ending billion dollar migration project. The answer for most such workloads is yes.