ARM Versus X86 - Insights Gained From The OSADL QA Farm

1y ago
10 Views
2 Downloads
2.20 MB
42 Pages
Last View : 1d ago
Last Download : 2m ago
Upload by : Annika Witter
Transcription

ARM versus x86 –insights gained from the OSADL QA FarmCarsten EmdeOpen Source Automation Development Lab(OSADL) eGEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

ARM versus x86 – Disclaimer The Open Source Automation Development Lab (OSADL)primarily takes care of using Open Source software in theindustry and in industrial products.One of our services is to run a quality assessment farm.We can use our farm data to compare ARM versus x86.However, since we only use Open Source software, ourresults are restricted to such software, end we – evidently –cannot make any conclusion on the behavior of anyproprietary software running on the tested architectures.Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

What is the OSADL QA Farm? (1)OSADL Test RackEight individual tablets Power supply 220 V, Ethernet, RS232 10/100/1000 Mb/s network switch with port mirroring Remote power switch with metering of every tablet Serial network adapter for every tablet KVM switch (optional) for every tablet One control server per rack Munin monitoring and Nagios event recording Additional scripts and tools developed by OSADL Designed and realized by kernel developer Thomas Gleixner Available at OSADL founding member Linutronix Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

OSADL QA Farm (2)Individual systems are mounted on removable tabletsEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

OSADL QA Farm (3)Cloud-based communication between test systems,data collectors and admin systemsVPN ChannelsTest center #1Test center #2Management,evaluationData collectorEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

osadl.org/QA – main menu OSADL Realtime QA FarmAbout - Hardware - CPUs - Benchmarks - Graphics - Benchmarks - Kernels - Boards/Distros Latency monitoring - Latency plots - System data - Profiles - Compare - Awards Overview of the processors under test and system locations ( 100 systems)Processor families: ARM MIPS PowerPC x86 Processor family ARM ( 20 systems)Manufacturers: Broadcom Freescale Marvell NXP TI Xilinx Processor family x86 ( 60 systems)Manufacturers: AMD Intel VIAEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

osadl.org/QA – system selectorEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

ARM versus x86 – criteria of comparison ofsuitability for embedded systems1. Run-time stability (uptime without crashes)2. Delay between market introduction and “reasonable”mainline Linux support3. Amount of unavoidable closed-source software4. Real-time support and real-time capabilities5. Computational efficiency (computing power by watts)6. Minimum power consumptionEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

ARM versus x86 – criteria that will not beconsidered1. Price and other purchase conditions2. Long-term availability3. Design support4. Many other individual criteriaEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine run-time stability?DayWeekMonthYearEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities?Signal pathExternal event,e.g. from a light barrier33Gate CPUlatency IRQReturn from systemcall in user space915Interrupt serviceroutineScheduling,context switch6IRQlatency30Total latency or preemption latencyEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities? Test setupPowerPC 750FX@600MHz64 MB SDRAM on SODIMM, 16 MB Flash-EPROM10/100 Mb/s Network2 serial channels RS232 and RS4852 TTL Outputs, 4 TTL Inputs4 Status LEDsOn-board FPGA339Interrupt serviceroutine15Scheduling,context switchEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities? Data setLine #10(No latency recording below 1 µs duration)0Histogram data 00000000Line #1176 (A total of 76 latency values between 10 and 11 µs duration)2238880020027(Most frequently observed latency values between 13 and 14 µs duration)184334302514[.]Line #1000 0(No overflow)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities? Latency plotEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities? Long-term (1)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine real-time capabilities? Long-term (2)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine computational efficiency?Power consumption versus P statesEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine computational efficiency?Power consumption versus P statesEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine computational efficiency? Select performance governor (max. clock frequency) Disable any sleep states Create Dhrystone and Whetstone test programs with samecompiler and comparable compiler optionsRun Dhrystone and Whetstone tests while measuring powerconsumption Calculate kilo Dhrystones/W Calculate Whetstones/WEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

How to determine computational efficiency? Select performance governor (max. clock frequency)leb!aGrNIaNpdRDisable any sleep statesmnAoacWyelbspeptcuoisnvCreate Dhrystone iler Run Dhrystoneand ptionadnaepyCalculatekilo Dhrystones/WtCalculate Whetstones/WEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Expected resultsCriterion / ProcessorARMx861. Suitability with respect to run-time stability?2. Suitability with respect to mainline Linux support delay?3. Suitability with respect to unavoidable closed-source software?4. Suitability with respect to real-time support and real-time capabilities?5. Suitability with respect to computational efficiency?6. Suitability with respect to minimum power consumption?Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Run-time stability examplesx86 Example of astable systemARM Example of anunstable systemEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Run-time stability ARM–All AM335x and i.MX6 systems are stable, no systemcrashes observed.–Other ARM systems (some of them very new, some of themvery cheap) are not yet stable, we are working on thisissue.–All x86 systems are stable, no systems crashes observed.x86Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Delay between market introduction and“reasonable” mainline Linux support for ARM ARM–Some older systems (introduced before 2011) still do nothave mainline Linux support.–Some systems (e.g. introduced in 2012) are currently gettingmainline Linux support.–Recently introduced systems got mainline Linux supportwithin months.Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Delay between market introduction and“reasonable” mainline Linux support for x86 x86–As a general rule, an x86 system has mainline Linux supportat the date of its introduction thanks to Intel's and AMD'sLinux support. The only exception to this rule may belacking or insufficient support of on-board graphics.Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Amount of unavoidable closed-sourcesoftwareWhich / ProcessorARMBIOSOn-board graphics controllerEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farmx86noYESYESYES/no

Results: Real-time support and real-timecapabilities Rule of thumb:An industrial computer should be able to continue executionof a user-space program that is waiting for an unpredictableasynchronous event within 10⁵ times its clock interval.Example:A clock frequency of 1 GHz is equivalent to a cycle durationof 1 ns. This duration times 10⁵ equals 100 µs. Thus, theworst-case latency of a 1-GHz processor should not exceed100 µs.Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an ARMprocessor @996 MHz (100 million cycles)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an ARMprocessor @996 MHz (22 billion cycles)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an x86processor @2900 MHz (100 million cycles)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an x86processor @2900 MHz (60 billion cycles)Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an x86processor (100 million cycles), apparently goodEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Result example: Real-time capabilities of an x86processor (150 billion cycles) SMI-related outliersEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Results: Computational efficiency (integer arithmetic)x86 Intel Core i7-E610@2533 MHz10.000(kDhrystone/W)ARM Freescale i.MX6@996 MHz (with/withoutperipheral devices)ARM TI AM3359@720 MHz1.000100x86 AMD G SeriesT40EC0/E1@1140 MHzx86 IntelCeleron J190@1990 MHz(former“Bay Trail”)x86 Intel Atom D525@1800 MHzARM BroadcomBCM2708@700 MHzx86 AMD Geode@500 MHz10x86 Intel Corei3-4360@3700 MHz(former “Haswell”)ARM Xilinx Zync@666 MHzx86 Intel Pentium@133 MHz VMEbus11995200020052010Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm2015

Results: Computational efficiency (floating point)x86 Intel Core i7-E610@2533 MHz1.000x86 AMD G SeriesT40EC0/E1@1140 MHzARM Freescale i.MX6@996 MHz (with/withoutperipheral devices)ARM BroadcomBCM2708@700 MHz(Whetstone/W)100x86 Intel CeleronJ1900 @1990 MHz(former “Bay Trail”)x86 Intel Atom D525@1800 MHzx86 Intel Pentium@133 MHz VMEbus10ARM TI AM3359@720 MHzx86 Intel Corei3-4360@3700 MHz(former “Haswell”)ARM Xilinx Zync@666 MHzx86 AMD Geode@500 MHz11995200020052010Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm2015

Result: Minimum power consumptionx86ARM0501001501.7 wattEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm200Watt250

Table of the resultsCriterion / ProcessorARMx861. Suitability with respect to run-time stability80% 100%2. Suitability with respect to mainline Linux support delay70% 100%3. Suitability with respect to unavoidable closed-source software90%4. Suitability with respect to real-time support and real-time capabilities90% 100%80%5. Suitability with respect to computational efficiency100%90%6. Suitability with respect to minimum power consumption100%70%Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

Estimated ARM and x86 suitability for embeddedsystems over time10.90.80.70.6ARMx860.50.40.30.20.1024 months agoNowIn 24 monthsEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

What is on our wish list? ARM–Immediate mainline Linux support, mo more vendor kernels–Linux device tree by default included in hardware delivery–Open Source graphics drivers–Open Source BIOS–Open Source graphics driversx86Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

What else is on our wish list? Every provider of industrial computing systems who isclaiming that at least one of the products runs on Linuxshould become OSADL member.Do not expect Linux-supporting companies such as Googleto fix your Linux issues!Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

The OSADL QA Farm and allother OSADL activities wouldnot be possible without theOSADL membersThe only OSADL member among the exhibitors of thisEmbedded Systems Design Conference!Embedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

ConclusionBoth ARM and x86 processors are relatively well suitable to beused in industrial and other embedded systems that requirestable, reliable and deterministic run-time behavior.Important advantages of ARM processors are No BIOS and, therefore, no BIOS-related problems Power consumption in the range of about 2 to 20 wattImportant advantages of x86 processors are Nearly perfect and early mainline Linux support Very high computing power and very short real-time responseEmbedded Systems Design ConferenceJuly 3, 2014, Munich, GermanyARM versus x86 – insights gained from the OSADL QA Farm

ARM versus x86 - insights gained from the OSADL QA Farm ARM versus x86 - Disclaimer The Open Source Automation Development Lab (OSADL) primarily takes care of using Open Source software in the industry and in industrial products. One of our services is to run a quality assessment farm. We can use our farm data to compare ARM versus x86.

Related Documents:

The x86 emulator also works such as Windows and NetBSD. TegraK-1 in Using Qemu Pentium2 266MHz equivalent Win98 is in HDBench, it was to use and Pentium4 1.8GHz equivalent to ExaGear-Desktop. In ARM Devices,Qemu is x86:ARM 3:1 (proportional to the floating point?) ARM x86 ARM x86 NetBSD Teokure Live Image on ARM chromebook

android-x86.org Android-x86 status update from lead developer Chih-Wei Huang . Virtualbox and VMware Player supported. 26-28 Sept. - A Coruña android-x86.org oreo-x86 features . marshmallow-x86 3.7 FORCE_AMDGPU cflag to fix function prototypes (maurossi)

Figure 1. Slowdown when emulating ARM applications on x86 versus running natively on ARM (top graph) and the reverse for native x86 applications in the bottom graph. stand the costs of using KVM/QEMU to abstract the ISA, we measured the slowdown when migrating an application (including the operating system) between KVM on x86 and QEMU on ARM.

Amazon EC2 64-bit: x86-64. SPARC 64 *15. x86-64. SPARC 64 *15. x86-64. IA64: . Sun Solaris SPARC. Sun Solaris x86-64: Sun Solaris SPARC. Sun Solaris x86-64: HP HP-UX Intel Itanium. . Technical Services may ask the customer to reproduce the issue on the Red Hat or SUSE distributions that are supported before

FFmpeg LGPL2.1 GPLv2 x86, ARM C MP4, Command line X server, command line Mist Server aGPLv3 x86, ARM, MIPS C, JS FLV, MP3, OGG Web browser Web browser red5 Apache x86, ARM Java FLV, MP4, Web browser, Flash Web browser, Flash Video streaming applications FFmpeg/VLC are suitable because

x86 instruction set o x86 instruction set started in 1978 with 16-bit CPUs First 32-bit CPU with x86-32/IA-32 released in 1985 o x86-64/x64 developed by AMD in 2003 (Intel refer to it as EM64T) 32-bit versus 64-bit CPU o 32-bit CPU can address up to 4 GB memory o 64-bit CPU can address up to 16 Exabytes (

A ARM B x86 C MIPS D VLIW E CISC 2. Was the full x86 instruction set we have today carefully planned out? Letter Answer A Yes B I wish I could unlearn everything I know about x86. I feel unclean. C Are you kidding? I’ve never seen a more poorly planned ISA! D *sob* E B, C, or D 3.

DEFINISI INVESTASI Investasi adalah komitmen atas sejumlah dana atau sumberdaya lainnya yang dilakukan pada saat ini, dengan tujuan memperoleh sejumlah