2nd CERN Advanced Performance Tuning workshop

21 Nov 2013, 09:00 → 22 Nov 2013, 18:00 Europe/Zurich

593/R-011 (CERN)

During this two day workshop focused on Advanced Performance Tuning, the latest trends in tuning will be presented by industry leaders ARM, Calxeda, Google and Intel.

The participants will have a unique opportunity to interact with architects, software experts and members of teams responsible for mainstream tools such as linux-perf and Intel VTune Amplifier.

A detailed agenda will be announced in the upcoming days and will include lectures as well as an opportunity for hands-on labs. The speakers include:

• Stas Bratanov (Intel), VTune team
• Andres S. Charif-Rubial (Versailles Exascale Computing lab), tuning expert
• Mike Chynoweth (Intel), PBA/xIF team, tuning expert
• Maria Dimakopoulou (Google), performance engineer
• Stephane Eranian (Google), senior engineer for linux-perf, libpfm, perfmon2 author
• Al Grant (ARM), performance expert
• Vincenzo Innocente (CERN), performance and software expert
• David Levinthal (Google, remote), GooDA author and tuning expert
• Andrzej Nowak (CERN openlab)
• Robert Richter (Calxeda), performance expert: linux-perf, ARM and AMD systems
• Mike Williams (ARM), debug and PMU architect
• Ahmad Yasin (Intel), PMU architect

If you would like the experts to take a look at your code in a specific tool, please prepare an SLC6 compatible package in advance, or have your own system ready. Openlab will provide dual-socket Ivy Bridge systems - to get advance access, please get in touch with Pawel Szostek.

Organizer: Andrzej Nowak, CERN openlab CTO office / PCC

To request your place, please click the "Apply" link below - due to limited space participation will be confirmed.

Thursday 21 November

09:00 → 10:15 Talks

09:00 Introduction

Speaker: Andrzej Nowak (CERN openlab)

Slides A.Nowak_-_PMU_user's_workshop_introduction_v2.pdf

09:30 Physics Software and Tuning Challenges + Discussion

Speaker: Vincenzo Innocente (CERN)

Slides Profiling@HEP.pdf

10:15 → 10:45 Break: Coffee break

10:45 → 12:10 Talks

10:45 An update on perf_events

In this talk, we give an overview of the latest developmements in the Linux kernel performance monitoring interface, perf_events, and related tools, such as perf. In particular, we describe load/store sampling support, event grouping, multi event profiling, Energy consumption, uncore counters, Haswell processor support.

Speaker: Stephane Eranian (Google)

Slides SE-CERN_PMU_workshop_2013.pdf

11:15 Improving perf_events measurement correctness

In this presentation, we talk about a correctness issue in the Performance Monitoring Unit (PMU) of recent Intel processors with Hyperthreading enabled. This issue introduces cross-threading corruption when hyperthreads measure incompatible events on sibling counters. As such, certain event combinations may produce unreliable results. We present an innovative approach to avoid the problem by introducing cross-thread dynamic event scheduling. We conclude with the implemented protocol's results and the challenges it raises.

Speaker: Maria Dimakopoulou (Google)

Slides Maria_Dimakopoulou-CERN_PMU_Workshop_2013.pdf

11:25 RAS and memory error reporting with perf

Strategies for RAS (Reliability, availability and serviceability) are necessary for enterprise systems in order to increase data integrity and system uptime. The current implementations in the Linux kernel to collect hardware errors are architecture dependent or even vendor specific. In order to unify hardware error reporting over architectures a new approach is needed. The talk shows how the perf event subsystem can be used for this. It also gives details about perf persistent events that keep running in the system after the creating process terminated.

Speaker: Robert Richter (Calxeda)

Slides RR_ras-perf-cern.pdf

11:55 AMD IBS and northbridge counters in perf

Speaker: Robert Richter (Calxeda)

Slides RR_amd-ibs-cern.pdf

12:10 → 14:00 Break: Lunch (on your own)

14:00 → 15:30 Talks

14:00 MAQAO: an analysis and optimization toolchain

MAQAO (Modular Assembly Quality Analyzer and Optimizer) analyzes binary codes and provides application developers with reports to optimize their code. The tool mixes both static code quality evaluation, and dynamic profiling and characterization. This is based on the ability to reconstruct low-level and high-level structures, such as basic blocks, loops, functions, and call-sites. Another main feature of MAQAO is its extensibility. Users can easily write their own plug-ins, using the embedded scripting language Lua. It allows fast prototyping of new tools based on MAQAO. We will present the three currently released modules. The first one is the profiler module (PERFEVAL) which aim is to detect function and loop hotspots. The second one is the code quality analyzer tool (CQA) which evaluates the code generated by a given compiler. Finally, we will present MIL, our binary instrumentation language.

Speaker: Andres S. Charif-Rubial (Exascale Computing Research Laboratory, Versailles)

Slides ACR_-_MAQAO.Toolchain.pdf

14:30 Intel VTune Amplifier: A Bridge to Performance, Parallelism, and Power (introduction to hardware collection)

We’re going to present Intel VTune Amplifier XE as Intel’s flagship performance analysis product and focus on a few aspects of HW-assisted SW analysis, including performance, power and threading efficiency, in the example of an N-body application run on both CPU and GPU. Additionally, in the process of our case study we will identify opportunities for further improvement of the tool and will ask the audience to share their opinions. CPU analysis is expected to be covered in the talk, while the GPU side will be left for offline studying because of time constraints.

Speaker: Stas Bratanov (Intel)

Slides Stas-VTune4CERN2013.pdf

14:50 Top Down Analysis – Never Lost With Perf Counters

Optimizing an application’s performance for a given microarchitecture has become painfully difficult. Increasing microarchitecture complexity, workload diversity, and the unmanageable volume of data produced by performance tools increase the optimization challenges. At the same time resource and time constraints get tougher with recently emerged segments. This calls for accurate and prompt analysis methods, adding further to the difficulty. Top-Down Analysis is a practical method to quickly identify true bottlenecks in out-of-order processors. The presented method uses designated performance counters in a structured hierarchical approach in order to quickly and, more importantly, correctly identify dominant performance bottlenecks. The developed method is adopted by multiple in-production tools including VTune. Feedback from VTune average users suggests that the analysis is made easier thanks to the simplified hierarchy which avoids the high-learning curve associated with microarchitecture details. Characterization results of this method are reported for the SPEC CPU2006 benchmarks as well as key enterprise workloads. We will walk through field case studies where the method guides field software optimizations, in addition to architectural exploration study for most recent generations of Intel Core products.

Speaker: Ahmad Yasin (Intel)

Slides Top_Down_for_CERN_2nd_workshop_-_Ahmad_Yasin.pdf

15:30 → 16:45 Hands-on: Hands-on and discussions

17:00 → 18:30 Break: ALICE experiment underground visit (speakers)

19:00 → 22:30 Break: Optional dinner

Friday 22 November

08:00 → 09:30 Talks

08:00 Intel Architecture and GOODA

Speaker: David Levinthal (Google)

document DL-useful_links.txt

Slides

• DL-cycle_accounting_and_gooda.pdf
• DL-CycleAccountingandPerformanceAnalysis.pdf
• DL-Micro-architecture.pdf

09:30 → 10:00 Break: Coffee Break

10:00 → 11:40 Talks

10:00 ARM in the server space

Speaker: Robert Richter (Calxeda)

Slides RR-arm-server-cern.pdf

10:15 Introduction to ARMv8

ARM business model of designing and licensing low-power IP building blocks has been phenomenally successful in transforming the mobile industry. The introduction of the ARMv8 architecture opens up low-power 64-bit computing in the same way. This talk will introduce the ARMv8 architecture and describe the debug and performance monitoring capabilities baked into it. It will give an opportunity to learn more about ARM and ARMv8.

Speaker: Michael Williams (ARM)

Slides MW-ARM_PerformanceMonitoring_MW_CERN2013.pdf

10:35 Software profiling on ARM

This presentation will give a brief overview of ARM's software profiling tools. This will mainly cover the Streamline tool for sampling-based profiling and software instrumentation, but also look at hardware trace.

Speaker: Al Grant (ARM)

Slides AG_-_cern2013-arm-profiling.pdf

10:55 Utilizing Performance Bottleneck Analyzer to Debug Issues on Intel’s Future SOCs

The presentation will cover the methodologies of the PBA (Performance Bottleneck Analyzer) toolset which has been maintained by Intel engineers for 7+ years to analyze workloads on future architectures. Using examples from software vendors we will show how PBA was able to find and fix issues that could not be identified with any other methodology or toolset on Intel’s future SOCs. PBA recreates very long flows of execution on the processor and then combines knowledge of processor events, power state and static assembly analysis to find and prioritize bottlenecks on Intel’s latest architectures. Filters are then applied to the data set to better call out issues that are impacting user experience, power or slow transactions to ensure the developer concentrates on the right issues to fix their problem. We will also showcase all of the functionalities using field examples from major software vendors. The talk will focus on how the collaborative framework has been used to share methodologies across multiple software vendor accounts and disciplines. We will also focus on some new capabilities allowing us to collect more performance monitoring data for less overhead than previously realized and syncing the run with multiple types of media.

Speaker: Michael W Chynoweth (Intel)

Slides MC_-_PBAforCERN_post_MC.pdf

11:40 → 12:30 Hands-on: Discussions and hands-on

12:30 → 14:00 Break: Lunch (on your own)