Multi-stack System Software

Jack LangeAssistant ProfessorUniversity of Pittsburgh

Summary• Commodity and HPC systems have been converging

– Commodity off the shelf components– Linux based HPC systems– Cloud computing

• Problem: Real HPC applications need HPC environments– Tightly coupled, massively parallel, and synchronized– Current services must provide dedicated HPC systems

• Can we co-host HPC applications on commodity systems?

• Dual Stack Approach– Provision the underlying software stack along with application– Commodity stack should handle commodity applications– HPC stack can provide HPC environment

• Current systems do support this, but…• Interference still exists inside the system software– Inherent feature of commodity systems

CoresSocket 1

Memory

1 2

3 4

CoresSocket 2

5 6

7 8

Memory

HPC PartitionCommodity Partition

User Space Partitioning

HPC vs. Commodity Systems

• Commodity systems have fundamentally different focus than HPC systems– Amdahl’s vs. Gustafson’s laws– Commodity: Optimized for common case

• HPC: Common case is not good enough– At large (tightly coupled) scales, percentiles lose meaning– Collective operations must wait for slowest node– 1% of nodes can make 99% suffer– HPC systems must optimize outliers (worst case)

Dual Stack Approach

• Partition– Segment the underlying hardware resources– Assign them to exclusively to specific workloads

• Isolate– Prevent interference from other workloads– Hardware: partitions must be course grained– Software: eliminate shared state

• Implementation– Independent system software running on isolated resources

HPC in the cloud• Clouds are starting to look like supercomputers…

– Are we seeing a convergence?• Not yet

– Noise issues– Poor isolation– Resource contention– Lack of control over topology

• Very bad for tightly coupled parallel apps– Require specialized environments that solve these problems

• Approaching convergence– Vision: Dynamically partition cloud resources into HPC and commodity zones– This talk: partitioning compute nodes with performance isolation

Commodity VMMs

• Virtualization is considered an “enterprise” technology– Designed for commodity environments– Fundamentally different, but not wrong!

• Example: KVM architecture issues– Userspace handlers– Fairly complex memory management– Locking and periodic optimizations– Presence of system noise

8

Palacios VMM• OS-independent embeddable virtual machine monitor

– Established compatibility with Linux, Kitten, and Minix

• Specifically targets HPC applications and environments– Consistent performance with very low variance

• Deployable on supercomputers, clusters (Infiniband/Ethernet), and servers– 0-3% overhead at large scales (thousands of nodes)

• VEE 2011, IPDPS 2010, ROSS 2011

http://www.v3vee.org/palaciosOpen source and freely available

Palacios/Linux

• Palacios/Linux provides lightweight and high performance virtualized environments– Internally manages dedicated resources

• Memory and CPU scheduling– Does not bother with “enterprise features”

• Page sharing/merging, swapping, overcommitting resources

• Palacios enables scalable HPC performance on commodity platforms

VMM Comparison

• Primary difference: Consistency– Requirement for tightly coupled performance at large scale

• Example: KVM nested paging architecture– Maintains free page caches to optimize performance

• Requires cache management– Shares page tables to optimize memory usage

• Requires synchronizationVMM % of exits Mean Std Dev # NPFS

KVM 52% 8804 5232 3,265,156

Palacios 50% 10876 2685 1,872,017

Hardware

Palacios VMM

HPC Linux

Linux Kernel

HPC Application

Palacios ResourceManagers

KVM

Linux Module Interface

Commodity Linux

Commodity Application(s)

Dual Stack Architecture• Partitioning at the OS level

• Enable cloud to host both commodity and HPC apps– Each zone optimized for the target applications

Evaluation• Goal: Measure VM isolation properties• Partitioned a single node into HPC and commodity zones

– Commodity Zone: Parallel Kernel compilation– HPC Zone: Set of standard HPC benchmarks– System:

• Dual 6-core AMD Opteron with NUMA topology • Linux guest environments (HPC and commodity)

• Important: Local node only– Does not promise good performance at scale– But, poor performance will magnify at large scales

Results

Palacios delivers consistent performance

Commodity VMMs degrade with contention

MiniFE: Unstructured implicit finite element solverMantevo Project -- https://software.sandia.gov/mantevo/index.html

Discussion• A dual stack approach can provide HPC environments on

commodity clouds– HPC and commodity workloads can dynamically share resources– HPC requirements can be met without fully dedicated resources

• Networking is still an open issue– Need mechanisms for isolation and partitioning– Need high performance networking architectures

• 1Gbit is not good enough• 10Gbit is good, Infiniband is better

– Need control over placement and topologies

Multi-stack Operating Systems• Future Exascale Systems are moving towards in situ organization• Applications traditionally have utilized their own platforms• Visualization, storage, analysis, etc

• Everything must now collapse onto a single platform

What this means for the OS• At Petascale we could optimize each environment separately

– Each had their own OS and hardware• At Exascale workloads will be co-located

– Can a single OS handle all workloads effectively?

• Probably Not– Each has different resource requirements and behaviors– Exascale will need to support multiple OS environments on the same

hardware

Beyond Virtualization• Virtualization imposes overhead– Power: requires transistors– Performance: small, but present– Interference: Still some dependencies on host OS

• Might not be available on exascale hardware

• Can we provide native partitioning?– We think so– Linux provides the ability to dynamically remove resources

(CPUs, memory, etc)– These can be taken over by a second OS

Hardware

Palacio VMMKitten

HPC Application

Linux

Commodity Application(s)

Para-native Architecture

• Provide LWK environment on a commodity system– Each zone optimized for the target applications

Approach

• OS partition created via offlined resources– CPUs, memory, PCI devices

• Secondary OS “booted” on offline resources

• Issues:– OS initialization

• Boot process• Resource discovery

– Coordination and communication– Security and safety

Dual Stack Memory

• Maybe we don’t need to provide an entirely separate OS– Instead selectively manage some resources

• Dual stack memory– Provide a separate memory management layer to Linux

• Features– Selectively manage heap per application– Provide applications with direct control over memory layout – Transparently back memory using large pages – Without overhead added by Linux

Hardware

HPC Application

Linux

Commodity Application(s)

Dual Stack Architecture

• Provide LWK memory manager on a commodity OS

LightweightMemory

Management

Commodity Linux

Initial on-demand Page faults

(500,000 – 600,000 cycles)

Performance Comparison

Linux Memory Management Lightweight Memory Management

Occasional Outliers(Large page coalescing)

Lowlevel noise

Conclusion• Commodity systems are not designed to support HPC workloads

– Different requirements and behaviors than commodity applications

• A multi stack approach can provide HPC environments in commodity systems– HPC requirements can be met without separate physical systems– HPC and commodity workloads can dynamically share resources– Isolated system software environments

Thank you

Jack LangeAssistant Professor University of Pittsburgh

jacklange@cs.pitt.edu

http://www.cs.pitt.edu/~jacklange