SPARC Systems - D3Sd3s.mff.cuni.cz/teaching/crash_dump_analysis/slides/08-sparcsys.pdf · SPARC...

18
http://d3s.mff.cuni.cz Crash Dump Analysis 2015/2016 CHARLES UNIVERSITY IN PRAGUE faculty of mathemacs and physics faculty of mathemacs and physics SPARC Systems SPARC Systems

Transcript of SPARC Systems - D3Sd3s.mff.cuni.cz/teaching/crash_dump_analysis/slides/08-sparcsys.pdf · SPARC...

http://d3s.mff.cuni.czCrash Dump Analysis 2015/2016

CHARLES UNIVERSITY IN PRAGUE

faculty of mathematics and physicsfaculty of mathematics and physics

SPARC SystemsSPARC Systems

Crash Dump Analysis 2014/2015 2Deadlocks and Hangs

MotivationMotivation

$ uname -aSunOS t7-1 5.11 11.3 sun4v sparc sun4v

Host name

Solaris version

SunOS version

ISA

PlatformOS name

Crash Dump Analysis 2014/2015 3Deadlocks and Hangs

Workstation and ServerWorkstation and Server

(Graphical) WorkstationDesktop computer with lots of RAM and CPU powerHigh-resolution color monitorStorage: SCSI tape (LTO), Fibre Channel...

(Enterprise) ServerRackmount computer with lots of RAM and CPU powerNo monitorStorage: Lots of disks, tape, ...

Crash Dump Analysis 2014/2015 4Deadlocks and Hangs

TimelineTimeline

1982 1984 1986 1988 1990 1992 1994 1996 1998 2000

Sun­1 Sun­2 Sun­3 Sun­4/SPARCstation Ultra Series

Ultra Ent. Enterprise

M68xx SPARC x86

Crash Dump Analysis 2014/2015 5Deadlocks and Hangs

Timeline (continued)Timeline (continued)

2000 2002 2004 2006 2008 2010 2012 2014 2016 2018

Fire

Fire

Java WS + Ultra

T Series

Oracle M Series

Blade

X Series

M68xx SPARC x86

Crash Dump Analysis 2014/2015 6Deadlocks and Hangs

PlatformPlatform

Motorola 68xxsun1, sun2, sun3

32-bit SPARCsun4, sun4c, sun4m, sun4d...

64-bit SPARCsun4u (u for UltraSPARC)sun4v (v for Virtualization) – T Series and current M Series

Crash Dump Analysis 2014/2015 7Deadlocks and Hangs

T SeriesT Series

UltraSPARC T1Sun Fire/Sun SPARC Enterprise T1000, T2000

UltraSPARC T2Sun SPARC Enterprise T5120/5140/5220/5240/5440

SPARC T3SPARC T3-1, T3-2, T3-4

SPARC T4SPARC T4-1, T4-2, T4-4

SPARC T5SPARC T5-2, T5-4, T5-8

SPARC M7SPARC T7-1, T7-2, T7-4

Crash Dump Analysis 2014/2015 8Deadlocks and Hangs

M SeriesM Series

SPARC64Sun SPARC Enterprise M3000, M4000, M5000, M8000, M9000

SPARC M5M5-32

SPARC M6M6-32

SPARC M7M7-8, M7-16

Crash Dump Analysis 2014/2015 9Deadlocks and Hangs

Reliability, Availability, ServiceabilityReliability, Availability, Serviceability

ISA FeaturesDefined behavior in different error scenarios

CPU FeaturesRegister content protectionInstruction restarting

System FeaturesRedundant componentsHot-replacable componentsService Processor (monitoring, alerts)

OS FeaturesPredictive self-healing (FMA)Hot-replacable components

Crash Dump Analysis 2014/2015 10Deadlocks and Hangs

Hardware PartitioningHardware Partitioning

System consist of modulesC (Compute), M (Memory), IO (I/O)CU+MU+IOU, CMU+IOU, CMIOUOlder systems use different terminology

Can be partitioned into domainsElectrically isolatedConfigured by service processorCan be changed dynamically at run timeTypically on module boundaries

Crash Dump Analysis 2014/2015 11Deadlocks and Hangs

FirmwareFirmware

IEEE 1275-1994 (Open Firmware, OpenBoot, OBP)IEEE standard (now considered withdrawn)Used with POWER/PowerPC (incl. Mac) and SPARCBased on Forth language interpreter'OK' promptCan be used to examine HW, debug the OS kernelFCode – architecture-indepenent driver/plugin residing on an add-on card or boardWorkstations use a graphical console

Solaris now does something similar on x86, too

Crash Dump Analysis 2014/2015 12Deadlocks and Hangs

Concurrent Multi-ThreadingConcurrent Multi-Threading

1 core – 8 concurrent instruction streamsCalled strandsEliminates waiting for memoryAlternative to out-of-order executionWorks well for parallel workloadsIntel's HT is somewhat similar (but only 2 streams)e.g. M7 has 32 cores, 32 x 8 = 256 strandsStatic vs. dynamic schedulingLater combined with out-of-order

Crash Dump Analysis 2014/2015 13Deadlocks and Hangs

Hyperprivileged modeHyperprivileged mode

Extra privilege levelUnprivileged -> Privileged -> Hyperprivileged

Extra level of addressingVirtual -> Real -> Physical

HypervisorPart of platform firmware, always presentAddress translation, resource partitioning

Used with different V18n technologiesLogical Domains (Oracle VM for SPARC)Solaris Kernel Zones

Crash Dump Analysis 2014/2015 14Deadlocks and Hangs

Logical Domains (Oracle VM for SPARC)Logical Domains (Oracle VM for SPARC)

Hardware PartitioningCPU, Memory, I/O (buses, slots, virtual functions)More flexible than hardware partitions

I/O VirtualizationNetwork, Disk, HBAVirtualized domains can be live-migrated

OS-independentDifferent OS versions and possibly even brands

Crash Dump Analysis 2014/2015 15Deadlocks and Hangs

Logical Domains continuedLogical Domains continued

Domain rolesControl (a.k.a. primary), Service, I/O, Root, Guest

Initial configurationInitially all resources belong to a single domainRemove some resources (CPUs, memory)Create new domainAssign resources to new domain

Many possible configurationsPartitioning only, fully virtual, HW-virtualized

Crash Dump Analysis 2014/2015 16Deadlocks and Hangs

Logical Domains Partitioning OnlyLogical Domains Partitioning Only

Memory

CPUs

Buses

Domain 1 Domain 2

Crash Dump Analysis 2014/2015 17Deadlocks and Hangs

Logical Domains VirtualizedLogical Domains Virtualized

Virtual serviceVirtual service Virtual service

Virtualdevice

Virtualdevice

Virtualdevice

Virtualdevice

Crash Dump Analysis 2014/2015 18Deadlocks and Hangs

Kernel Zones (Solaris)Kernel Zones (Solaris)

Solaris ZonesFramework for creating virtual OS instancesSimple user interfaceSupport resource controlStandard zones use OS-level virtualization

Branded ZonesLinux brand (defunct)S8, S9, S10

Kernel ZonesUse HW-based virtualizationIntel-VT/AMD-V on x86sun4v virtualization on SPARClive migration