WP3 Tools: distributed molecularsimulations and remote submissioninterface

Toni Giorgino

Computational Biochemistry and Biophysics Lab / GRIB-IMIMwww.multiscalelab.org

22-23 jan 2009

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 13

MD at the microsecond scale

� Full-atom MD is computationally expensive� All solvation water molecules modeled

� Non-bonded interactions

� All degrees of freedom in a protein

z

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 14

MD at the microsecond scale

� Full-atom MD is computationally expensive� All solvation water molecules modeled

� Non-bonded interactions

� All degrees of freedom in a protein

z

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 15

� Tackled at several scales� Combine advances of large-scale

infrastructures ⇒ distributed computing� ...with those in commodity computing

architectures ⇒ GPU

One node(CPU)

0.4 ns/day

1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 16

� Tackled at several scales� Combine advances of large-scale

infrastructures ⇒ distributed computing� ...with those in commodity computing

architectures ⇒ GPU

One node(CPU)

0.4 ns/day

Within-node 1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 17

� Tackled at several scales� Combine advances of large-scale

infrastructures ⇒ distributed computing� ...with those in commodity computing

architectures ⇒ GPU

One node(CPU)

0.4 ns/day

One node,accelerated4.5 ns/day

Within-node 1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 18

� Tackled at several scales� Combine advances of large-scale

infrastructures ⇒ distributed computing� ...with those in commodity computing

architectures ⇒ GPU

One node(CPU)

0.4 ns/day

One node,accelerated4.5 ns/day

Within-node Among-nodes1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 19

� Tackled at several scales� Combine advances of large-scale

infrastructures ⇒ distributed computing� ...with those in commodity computing

architectures ⇒ GPU

One node(CPU)

0.4 ns/day

One node,accelerated4.5 ns/day

Distributed computing

X 1,700, growing

Within-node Among-nodes1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 20

Tools

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

X 1,700, growing

Within-node Among-nodes1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 21

Tools

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

X 1,700, growing

Within-node Among-nodes1

AceMD

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 22

Tools

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

X 1,700, growing

Within-node Among-nodes1

AceMD Boinc

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 23

Tools

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

X 1,700, growing

Within-node Among-nodes1

AceMD Boinc

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 24

Computing architecture

� Based on BOINC (Berkeley Open Infrastructure for Network Computing)

� Allows (loosely-coupled) computations tobe distributed over the Internet

� Volunteers contribute CPU cycles� Not MD-specific� Public statistics at www.boincstats.com

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 25

Using computing power

• The infrastructure becomes useful for the community with a distributed submission system

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

× 2,000, growing

Within-node Among-nodes1

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 26

Using computing power

• The infrastructure becomes useful for the community with a distributed submission system

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

× 2,000, growing

Within-node Among-nodes1

Sim

. ent

rypo

int

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 27

Using computing power

• The infrastructure becomes useful for the community with a distributed submission system

One node(CPU)

0.25 ns/day

One node,accelerated4.6 ns/day

Distributed computing

× 2,000, growing

Within-node Among-nodes1

Sim

. ent

rypo

int

Research team 1 – hERGScientist 2 – gA…Research team N – protN

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 28

Remote simulation submission

• Pre-defined “best practice” submission protocols

• Researchers stage in/out files– They are taken in charge by Boinc

• Accounting: – Given the scale of deployment, increasing

“trustedness” associated to each simulation

– FLOPS consumed

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 29

Example: submission

1. Client parses local files and prepares a description of the simulation (may contain user-supplied metadata)

2. Inputs are uploaded to a staging area3. Execution

• Consis check• Move to execution area• Launch simulation

4. Results continuously produced; “work units” spanned until the protocol ends (# of ns × bins)

5. Clients periodically retrieve & delete results

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 30

Technically

� boinc_submit− Uploads local input files

− Creates config, launches WUs, etc

− Returns immediately

� boinc_stat� boinc_retrieve� [boinc_cancel et al]

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 31

Common traits

� All client commands have a notion of username and password to send forauthentication

� Authentication checked with Boinc protocol

� Server remembers uploaded files for laterdeletion (server-side “RemoteWU” entity)

� A RemoteWU has an ID for the clients toretrieve

− ideally equal to the usual WU id

VPH NoE Meeting - Oxford Thursday, January 22, 2009 Toni Giorgino 32

Outlook

• Expose remote submission interfaces• Identify ontologies for inclusion of relevant

annotations• Introduce metadata at least for annotating

inputs (to be carried out in outputs)