Challenges in Coordinating Large-Scale CPU and I/O Access

1
Challenges in CoordinatingLarge-ScaleCPU and
I/O Access

• Douglas Thain
• University of Wisconsin
• thain_at_cs.wisc.edu
• 16 September 2003

2
Integrating CPU and Data

• CPU Management
• Process Fork/Kill
• CPU Allocation
• Auth/Auth/Acct
• Time Sharing
• Multiplexing
• Checkpointing
• Dependency Mgmt
• Fault-Tolerance

• Data Management
• Low Level I/O
• Space Allocation
• Auth/Auth/Acct
• Replica Mgmt
• Consistency Mgmt
• Nesting
• Collection
• Aggregation

Integration??? Transparency Discovery Binding Faul
t Tolerance Performance
Increasing Abstraction
Increasing Abstraction
3
Pieces of the Puzzle
RFIO
dcache pnfs
Condor
FTP
Condor-G
NeST
Stork
Gatekeeper
Operating Systems
Parrot
ClassAds
Raw Storage
Raw CPUs
Disk Router
4
CPU Access Path
Local Job Queue
I/O Agent
Local Data Queue
Data Access Path
5
Possibilities for Integration

• Integration in the I/O Agent
• At run-time, change one resource allocation in
  response to another.
• Keep control close to the job.
• Integration in the Resource Broker
• Resources constrain each other after job
  submission, but resources are selected.
• A 3rd-party broker matches resources and reqs.
• Integration in the Workflow Manager
• High level planner allocates resource types
  independently and reacts at large granularity.
• Keep control close to the user.

6
Remote Job Queue
Local Manager
Users App
(POSIX)
Parrot I/O Agent
FTP
RFIO
NeST
Chirp
FTP Server
RFIO Server
NeST Server
Chirp Server
7
2 - Integration in the Broker
Refers to NearestStorage.
Knows where NearestStorage is.
Resource Broker
Job Ad
Machine Ad
Storage Ad
match
Storage
Machine
Job
8
ClassAd Example

• Job Ad
• Type Job
• Cmd cmsim.exe
• Owner thain
• Requirements
• (OpSysLINUX)
• (NearestStorage.
• HasCMS)

• Machine Ad
• Type Machine
• Name vulture
• OpSys Linux
• Requirements
• (Ownerthain)
• NearestStorage
• (TypeStorage)
• (Nameturkey)

Storage Ad Type Storage Name
turkey CMSPath /cms HasCMS True
9
3 - Integration in the Workflow

• job a a.condor
• job b b.condor
• job c c.condor
• job d d.condor
• parent a child c
• parent b child d

a
b
c
d
10
Workflow Language 2

• volume v1
• ftp//archive/mydata
• mount v1 a /data
• mount v1 b /data
• volume v2 scratch
• mount v2 a /tmp
• mount v2 c /tmp
• volume v3 scratch
• mount v3 b /tmp
• mount v3 d /tmp

v2
v3
a
b
c
d
11
Workflow Language 3
v2
v3
a
b

• extract v2 x ftp//home/out.1
• extract v3 x ftp//home/out.2

x
x
c
d
out.1
out.2
12
People and Projects

• Stork Data Transfer Scheduling
• Tevfik Kosart (kosart_at_cs.wisc.edu)
• DiskRouter Data Transfer Overlay
• George Kola (kola_at_cs.wisc.edu)
• NeST Network Storage Technology
• John Bent (johnbent_at_cs.wisc.edu)
• Joseph Stanley (jass_at_cs.wisc.edu)
• Parrot Transparent I/O Access
• Douglas Thain (thain_at_cs.wisc.edu)