Holistic Aggregate Resource Environment

Holistic Aggregate Resource Environment

Eric Van Hensbergen (IBM Research)Ron Minnich (Sandia National Labs)

Jim McKie (Bell Labs)Charles Forsyth (Vita Nuova)

David Eckhardt (CMU)

Overview

Sequoia

BG/LBG/L

Red Storm

Research Topics• Pre-requisite: reliability and application driven design is pervasive in all

explored areas

• Offload/Acceleration Deployment Model

• Supercomputer needs to become an extension of scientist's desktop as opposed to batch driven, non-standard run-time environment.

• Leverage aggregation as a first-class systems construct to help manage complexity and provide a foundation for scalability, reliability, and efficiency.

• Distribute system services throughout the machine (not just on io-node)

• Interconnect Abstractions & Utilization

• Leverage HPC interconnects in system services (file system, etc.)

• sockets & TCP/IP don't map well to HPC interconnects (torus and collective) and are inefficient when hardware provides reliability

Right Weight Kernel

• General purpose multi-thread, multi-user environment

• Pleasantly Portable

• Relatively Lightweight (relative to Linux)

• Core Principles

• All resources are synthetic file hierarchies

• Local & remote resources accessed via simple API

• Each thread can dynamically organize local and remote resources via dynamic private namespace

Aggregation• extend BG/P aggregation model

beyond I/O and CPU node barrier

• allow grouping of nodes into collaborating aggregates with distributed system services and dedicated service nodes

• allow specialized kernel for file service, monitoring, checkpointing, and network routing

• parameterized redundancy, reliability, and scaling

• allows dynamic (re-) organization of programming model to match the (changing) workload

local service

remote services

local service proxy service aggregate service

Topology

Desktop Extension

• Users want super computers to be an extension of their desktop

• Current parallel model is traditional batch model

• Workloads must use specialized compilers and be scheduled from special front-end node. Results are collected into a separate file system

• Monitoring and job control through web interface or MMCS command line

• Very difficult development environment and lack of interactivity limits productivity of execution environment

• Proposed Research

• leverage library OS commercial scale-out work to allow tighter coupling between desktop environment and super computer resources

• Construct runtime environment which includes some reasonable subset of support for typical Linux run-time requirements (glibc, python, etc.)

Extension Example

Mac pSeries I/O CPU

brasil

osx

brasil

Linux Plan 9 Plan 9

app app app

...toruscollective10GB Etherssh

internet

Native Interconnects

• Blue Gene specialized networks are used primarily by user space run-time

• Hardware is directly accessed by user space runtime time environment and are not shared leading to poor utilization

• Exclusive use of tree network for I/O limits bandwidth and reliability

• Proposed Solution

• Light weight system software interfaces to interconnects so that they can be leveraged for system management, monitoring, and resource sharing as well as user applications

Protocol Exploration• The Blue Gene networks are unusual (eg, 3D torus carrying 240-byte payloads)

• IP Works, but isn’t well matched to the underlying capabilities

• We want an efficient transport protocol to carry 9P messages & other data streams

• Related Work: IBM’s ‘one-sided’ messaging operations [Blocksome et al]

• It supports both MPI and non-MPI applications such as Global Arrays

• Inspired by the IBM messaging protocol, we think we might do better than just IP

• Years ago there was much work on lightweight protocols for high-speed networks

• We are using ideas from that earlier research to implement an efficient protocol to carry 9P conversations

Project Roadmap

0 1 2 3

Year

Hardware Support

Systems Infrastructure

Evaluation, Scaling, & Tuning

Milestones (Year 1)BASIC

BASELINE

Initial Boot10 GB Ethernet

Collective NetworkInitial Infrastructure

SMP SupportLarge Pages

Torus NetworkNative Protocol

Baseline

0 10 20 30 40 50

Weeks

2009

BG/L (circa 2007)

PUSH

!"#$$%

&'(()*+,-.# /0$1-.$#2'3

,-.#

,-.#

,-.#

,-.#

,-.#

,-.#!"#$$%

&'(()*+

!"#$$%

&'(()*+

!"#$$%

&'(()*+

!"#$$%

&'(()*+

!"#$$%

&'(()*+

!"#$$%

&'(()*+,-.#

,-.#

,-.#

,-.#

,-.#

,-.#

4#(0$1-.$#2'3 ,-.#!"#$$%

&'(()*+

Figure 1: The structure of the PUSH shell

We have added two additional pipeline operators, a multiplexing fan-out(|<[n]), and a coalescingfan-in(>|). This combination allows PUSH to distribute I/O to and from multiple simultaneousthreads of control. The fan-out argument n specifies the desired degree of parallel threading. If noargument is specified, the default of spawning a new thread per record (up to the limit of availablecores) is used. This can also be overriden by command line options or environment variables. Thepipeline operators provide implicit grouping semantics allowing natural nesting and composibility.While their complimentary nature usually lead to symmetric mappings (where the number of fan-outs equal the number of fan-ins), there is nothing within our implementation which enforces it.Normal redirections as well as application specific sources and sinks can provide alternate datapaths. Remote thread distribution and interconnect are composed and managed using syntheticfile systems in much the same manner as Xcpu [11], pushing the distributed complexity into themiddleware in an language and runtime neutral fashion.

PUSH also di!ers from traditional shells by implementing native support for record basedinput handling over pipelines. This facility is similar to the argument field separators, IFS andOFS, in traditional shells which use a pattern to determine how to tokenize arguments. PUSHprovides two variables, ORS and IRS, which point to record separator modules. These modules(called multiplexors in PUSH) split data on record boundaries, emitting individual records that thesystem distributes and coalesces.

The choice of which multipipe to target is left as a decision to the module. Di!erent dataformats may have di!erent output requirements. Demultiplexing from a multipipe is performedby creating a many to one communications channel within the shell. The shell creates a readerprocesses which connects to each pipe in the multipipe. When the data reaches an appropriaterecord boundary a bu!er is passed from the reader to the shell which then writes each record bu!erto the output pipeline.

An example from our particular experience, Natural Language Processing, is to apply an ana-lyzer to a large set of files, a ”corpus”. User programs go through each file which contain a list ofsentences, one sentence per line. They then tokenize the sentence into words, finding the part ofspeech and morphology of the words that make up the sentence. This sort of task maps very wellto the DISC model. There are a large number of discrete sets of data whose order is not necessarilyimportant. We need to perform a computationally intensive task on each of the sentences, whichare small, discrete records and ideal target for parallelization.

PUSH was designed to exploit this mapping. For example, to get a histogram of the distribution

2

push -c ’{ ORS=./blm.dis du -an files |< xargs os chasen | awk ’{print \$1}’ | sort | uniq -c >| sort -rn

}’

Early FTQ Results

Strid3 Y= AX + Y

Time for 1024 iterations

Timein

seconds

“Stride”, i.e. distance between scalars

Application Support• Native

• Inferno Virtual Machine

• CNK Binary Support

• Elf Converter

• Extended proc interface to mark processes as “cnk procs”

• Transition once the process execs, and not before

• Shim in syscall trap code to adapt arg passing conventions

• Linux Binary Support

• Basic Linux binary support

• Functional enough to run basic programs (Python, etc.)

Publications• Unified Execution Model for Cloud Computing; Eric Van Hensbergen, Noah Evans, Phillip Stanley-

Marbell. Submitted to LADIS 2009; October 2009.

• PUSH, a DISC Shell; Eric Van Hensbergen, Noah Evans. To Appear in the Proceedings of the Principles of Distributed Computing Conference; August 2009.

• Measuring Kernel Throughput on BG/P with the Plan 9 Research Operating System; Ron Minnich, John Floren, Aki Nyrhinen. Submitted to SC 09; November 2009.

• XCPU2: Distributed Seamless Desktop Extension; Eric Van Hensbergen, Latchesar Ionkov. Submitted to IEEE Clusters 2009; October 2009.

• Service Oriented File Systems; Eric Van Hensbergen, Noah Evans, Phillip Stanley-Marbell. IBM Research Report (RC24788), June 2009

• Experiences Porting the Plan 9 Research Operating System to the IBM Blue Gene Supercomputers; Ron Minnich, Jim McKie. To appear in the Proceedings of the International Conference on Supercomputing (ISC); June 2009.

• System Support for Many Task Computing; Eric Van Hensbergen and Ron Minnich. In the Proceedings of the Workshop on Many Task Computing on Grids and Supercomputers; November 2008.

• Holistic Aggregate Resource Environment; Charles Forsyth, Jim McKie, Ron Minnich and Eric Van Hensbergen. In the ACM Operating Systems Review; January 2008.

• Night of the Lepus: A Plan 9 Perspective on Blue Gene's Interconnects; Charles Forsyth, Jim McKie, Ron Minnich and Eric Van Hensbergen. In the proceedings of the second annual international workshop on Plan 9; December 2007

• Petascale Plan 9. USENIX 2007

http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=4777907




http://portal.acm.org/citation.cfm?id=1341327

http://portal.acm.org/citation.cfm?id=1341327

http://plan9.bell-labs.com/iwp9/papers/26.lepus.pdf

http://plan9.bell-labs.com/iwp9/papers/26.lepus.pdf

http://go.cs.bell-labs.com/fastos/doc/PetscalePlan9.pdf




Next Steps• Infrastructure Scale Out

• File Services

• Command Execution

• Alternate Internode Communication Models

• Fail in place software RAS models

• Applications (Linux binaries and native support)

• Large Scale LINPACK Run

• Explore Mantevo Application Suite

• (http://software.sandia.gov/mantevo)

• CMU Working on Native Quake port

http://software.sandia.gov/mantevo

http://software.sandia.gov/mantevo

Acknowledgments

• Computational Resources Provided by DOE INCITE Program. Thanks to the patient folks at ANL who have supported us bringing up Plan 9 on their development BG/P

• Thanks to IBM Research Blue Gene team and the Kittyhawk Team for guidance and support.

Questions? Discussion?http://www.research.ibm.com/hare

http://www.research.ibm.com/hare

http://www.research.ibm.com/hare

Backup

IBM Research, Sandia National Labs, Bell Labs, and CMU

(c) 2008 IBM CorporationSystems Support for Many Task Computing 11/17/200824



Plan 9 Characteristics

Kernel Breakdown - Lines of CodeArchitecture Specific Code

BG/P: ~14,000 lines of code

Portable CodePort: ~25,000 lines of code

TCP/IP Stack: ~14,000 lines of code

Binary Sizes415k Text + 140k Data + 107k BSS



Why not Linux?Not a distributed system

Core systems inflexible

VM based on x86 MMU

Networking tightly tied to sockets & TCP/IP w/long call-path

Typical installations extremely overweight and noisy

Benefits of modularity and open-source advantages overcome by complexity, dependencies, and rapid rate of change

Community has become conservative

Support for alternative interfaces waning

Support for large systems which hurts small systems not acceptable

Ultimately a customer constraint

FastOS was developed to prevent OS monoculture in HPC

Few Linux projects were even invited to submit final proposals



Everything Represented as File Systems

HardwareDevices

SystemServices

ApplicationServices

Disk

Network

TCP/IP Stack DNS

GUI

/dev/hda1

/dev/hda2

/dev/eth0

/net /arp /udp /tcp /clone /stats /0 /1 /ctl /data /listen /local /remote /status

/net /cs /dns

/win /clone /0 /1 /ctl /data /refresh /2

Console, Audio, Etc. Wiki, Authentication, and Service Control

Process Control, Debug, Etc.



Plan 9 Networks

Internet

High Bandwidth (10 GB/s) Network

LAN (1 GB/s) Network

Wifi/EdgeCable/DSL

ContentAddressable

Storage

FileServer

CPUServers

CPUServers

PDASmartphone

Term

TermTermTerm

Set Top Box

ScreenPhone

(‏ (‏(‏



Aggregation as a First Class Concept

Local Service Aggregate Service

Remote Service

Proxy Service

Remote ServiceRemote Service



Issues of Topology



File Cache Example

Proxy ServiceMonitors access to remote file server & local resourcesLocal cache modeCollaborative cache modeDesignated cache server(s)‏Integrate replication and redundancyExplore write coherence via “territories” ala Envoy

Based on experiences with Xget deployment modelLeverage natural topology of machine where

possible.



Monitoring Example

Distribute monitoring throughout the systemUse for system health monitoring and load balancingAllow for application-specific monitoring agents

Distribute filtering & control agents at key points in topology

Allow for localized monitoring and control as well as high-level global reporting and control

Explore both push and pull methods of modelingBased on experiences with supermon system.



Workload Management Example

Provide file system interface to job execution and scheduling.

Allows scheduling of new work from within the cluster, using localized as well as global scheduling controls.

Can allow for more organic growth of workloads as well as top-down and bottom-up models.

Can be extended to allow direct access from end-user workstations.

Based on experiences with Xcpu mechanism.



Right Weight Kernels Project (Phase I)‏

MotivationOS Effect on ApplicationsMetric is based on OS Interference on FWQ & FTQ benchmarks.AIX/Linux has more capability than many apps needLWK and CNK have less capability than apps want

ApproachCustomize the kernel to the application

Ongoing ChallengesNeed to balance capability with overhead



Why Blue Gene?

Readily available large-scale clusterMinimum allocation is 37 nodes

Easy to get 512 and 1024 node configurations

Up to 8192 nodes available upon request internally

FastOS will make 64k configuration available

DOE interest – Blue Gene was a specified target

Variety of interconnects allows exploration of alternatives

Embedded core design provides simple architecture that is quick to port to and doesn't require heavy weight systems software management, device drivers, or firmware

Holistic Aggregate Resource Environment

Technology

Transcript of Holistic Aggregate Resource Environment