When it comes to software performance, ideally we want two things:

1. The performance of our program doesn’t change as the features are added.
2. The runtime of the program grows linearly to the data set size.

But this often fails to happen. As new features are added or the dataset grows, the software starts running much slower than one would expect.

In this talk we will explore what are the factors behind this slowdown. We start from software architectural issues, move on to memory programs and talk about hardware inefficiencies that kill software performance

Personal computers and devices are unbelievably fast but often struggle to keep up with mundane tasks. Why hasn't software managed to scale in performance with hardware? This talk looks at some of the reasons, the characteristics of fast systems, the limits of human perception, and the need to rethink how software is authored to build efficient systems.

Most modern performance work is centered around technologies and tools. Following Jeff Bezos' advice, this talk, instead of trying to predict the future, covers some of the oldest mathematical principles and ways they've shaped modern software.

Parallel and distributed computing, caching and proxy computations are just a few examples of leveraging associativity, idempotence, Kleisli category and functors.

Armed with such powerful tools, it's time to open an era of efficient compute, but how can we be sure that we are not trading correctness for performance? Formal verification methods to the rescue. We'll see how TLA+ and Isabelle can be used to prove correctness of our ideas.

ByteDance in-house Redex passes A handful Redex passes have been developed in ByteDance in addition to existing Redex passes. We'd like to share some of them, RenameClass, StringBuilderOutliner, KtDataClass. These three passes saved about ~1M, ??%.

*Exploring ART Optimization Complement * ART optimization is triggered during both jit and dex2aot, however both are running with restricted computation power, mostly running on a phone. This implies that opt passes are light or caps on code size, this leaves us some opportunities of offline dex optimization.

Beyond Android Apps! As in the past year, we learnt a lot of bytecode level optimization, we are extending this learning to several other similar App models, for example, Unity app optimization.

Modern hardware is highly parallel, but not only in terms of multiprocessing. There are many other forms of parallelism that, if used correctly, can greatly boost program efficiency — and without requiring more CPU cores. One such type of parallelism actively adopted by CPUs is "Single Instruction, Multiple Data" (SIMD): a class of instructions that can perform the same operation on a block of 16, 32, or 64 bytes of data in one go, yielding a proportional speedup over scalar code.

While SIMD shares many similarities with classic multiprocessor computing, it is quite different and often requires creative use of the instruction set. In this talk, we will give a general introduction to the technology (focusing on x86/AVX2), derive and implement several state-of-the-art SIMD algorithms, and discuss their use in impactful open-source projects.

In recent years, the number of software services that are running on datacenter and cloud computing platforms has increased rapidly. Datacenters are expensive to operate and represent a significant portion of the world's energy consumption. Efforts to improve the efficiency of software that runs on datacenters can greatly reduce the energy use and costs of operations and improve the responsiveness of the service.

This talk will present techniques that cloud vendors use to reduce costs and optimize their software. We'll review a few tools and methodologies for identifying performance optimization opportunities and for evaluating the potential benefits of these optimizations. We'll cover the five performance optimization categories and explain how they apply to different scenarios.

The growing scale and complexity of data and ML workloads demands enormous compute power. Bodo is a new compute platform that aims to improve efficiency and performance of ETL pipelines through automated parallelization of native Python and SQL workloads.

Bodo’s resource efficiency and extreme performance in compute heavy workloads makes ETL orders of magnitude faster and more cost effective than popular solutions based on distributed computing such as Spark, Dask and Ray.

In this talk, we’ll discuss the high level overview of Bodo, some of the performance benchmarks, and innovations we did to save more than 50% infrastructure cost and gain more than 10x performance improvements saving for our customers.

In this session hosted by Taras Tsugrii, speakers from both days of Performance Summitt 2022 come together to explore what's next in performance improvement, and to answer any lingering questions posted by the talks.