Andrew Prout, William Arcand, David Bestor, Chansup Byun, Bill Bergeron, Matthew Hubbell, Jeremy Kepner, Peter Michaleas, Julie Mullen, Albert Reuther,

Antonio Rosa

2012 IEEE High Performance Extreme Computing Conference

10 - 12 September 2012

Scalable Cryptographic Authentication for High

Performance Computing

This work is sponsored by the Department of the Air Force under Air Force contract FA8721-05-C-0002. Opinions, interpretations, conclusions and recommendations are those of the author and are not necessarily endorsed by the United States Government.

HPEC 2012 - 2AJP 9/12/2012

• What is the LLGrid

• The Problem: External services authentication

• The Solution: Cryptographic authentication

• Results

Outline

HPEC 2012 - 3AJP 9/12/2012

• LLGrid is a ~500 user ~2000 processor system

• World’s only desktop interactive supercomputer– Dramatically easier to use than any other supercomputer– Highest fraction of staff using (20%) supercomputing of any

organization on the planet

• Foundation of Supercomputing in Massachusetts

LLGrid System Architecture

LAN Switch

Network Storage

Resource Manager

ConfigurationServer

Compute NodesService Nodes Cluster Switch

To Lincoln LAN

Users

LLAN

HPEC 2012 - 4AJP 9/12/2012

All jobs run on LLGrid

LLGrid Usage

1 10 100 1000

Tota

l Jo

b d

ura

tio

n (

seco

nd

s)

1

10

0

100

00

1M

Classic Supercomputing

Inte

ract

ive

Sup

erco

mpu

ting

Processors used by Job

TX-2500 (952 Cores)TX-X (220 Cores)TX-3d (540 Cores)

• Desktop Computing– CPU-time <20 minutes

• Classic Supercomputing– Wall-clock time >3 hours

• Interactive Supercomputing – Between desktop and classic

supercomputing– Shortens the “time to

insight”– Ten development turns/day

instead of one turn/week

• Desktop Computing– CPU-time <20 minutes

• Classic Supercomputing– Wall-clock time >3 hours

• Interactive Supercomputing – Between desktop and classic

supercomputing– Shortens the “time to

insight”– Ten development turns/day

instead of one turn/week

Des

ktop

Com

putin

g

HPEC 2012 - 5AJP 9/12/2012

• What is the LLGrid

• The Problem: External services authentication

• The Solution: Cryptographic authentication

• Results

Outline

HPEC 2012 - 6AJP 9/12/2012

• As the line between a shared supercomputer and a “really powerful personal computer” blurs, users expect to have access to network resources (storage, svn, cvs, etc).

Challenges withInteractive Supercomputing

Challenge: Users expect seamless access to other network resources from the HPC.

HPEC 2012 - 7AJP 9/12/2012

• However these commands raise security concerns.– They store passwords as plain-text on the HPC central storage.– Password synchronization has made this password very sensitive.

Challenges withInteractive Supercomputing

Challenge: Ensure seamless access without putting the user’s “one common password” at risk.

“S3cr3t”

HPEC 2012 - 8AJP 9/12/2012

• What is the LLGrid

• The Problem: External services authentication

• The Solution: Cryptographic authentication

• Results

Outline

HPEC 2012 - 9AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

HPEC 2012 - 10AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

HPEC 2012 - 11AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

Authentication RequestA

HPEC 2012 - 12AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

Authentication RequestA

A

HPEC 2012 - 13AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

Authentication Request

Signed Authentication Responseand copy of PKI certificate

A

A

HPEC 2012 - 14AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

Authentication Request

Signed Authentication Responseand copy of PKI certificate

A

A

A

HPEC 2012 - 15AJP 9/12/2012

• Cryptographic authentication of clients using X509 PKI certificates has long been part of the SSL and TLS standards.

• The root of trust will certify that a specific keypair belongs to a specific user or process.

Cryptographic Authentication

User Server

Connection Request

Authentication Request

Signed Authentication Responseand copy of PKI certificate

A

A

A

Access Granted: Welcome Andy!

HPEC 2012 - 16AJP 9/12/2012

• Cryptographic authentication depends on both the security of the user’s private key and access to it.– Storing the private key on central storage is little different than

storing a user’s password.

Challenges withCryptographic Authentication

Challenge: Where to store the private key?

HPEC 2012 - 17AJP 9/12/2012

• Cryptographic authentication depends on both the security of the user’s private key and access to it.– Storing the private key on central storage is little different than

storing a user’s password.

Challenges withCryptographic Authentication

No guarantee the key won’t be lost, copied or left unprotected.

HPEC 2012 - 18AJP 9/12/2012

• One traditional solution is to store the key on the client system and forward authentication requests back to the user’s system.– Could be on the client system or in a smart card.

Challenges withCryptographic Authentication

HPEC 2012 - 19AJP 9/12/2012

Challenges withCryptographic Authentication

Forwarding requests back doesn’t work forsemi-interactive computing or background jobs.

Poof!

• One traditional solution is to store the key on the client system and forward authentication requests back to the user’s system.– However this fails if the user disconnects from the HPC.

HPEC 2012 - 20AJP 9/12/2012

Challenges withCryptographic Authentication

Poof!

• Connecting smart cards to the HPC is not practical.– Some network-attached key storage devices exist, but their practical

benefit in this scenario is questionable.

HPEC 2012 - 21AJP 9/12/2012

Challenges withCryptographic Authentication

Poof!

• We implemented a virtual smart card to run on each node.– Allows for keys to be used on any node, connected or disconnected.– Allows for different keys on each node.

HPEC 2012 - 22AJP 9/12/2012

• Uses the smart card communication API: PKCS#11.

• Authenticates users and allows authorized users to perform cryptographic operations.

• Protects private keys from being copied, even by authorized users of the key.

• High throughput capability & low latency.– Physical smart cards have a latency approximately 800-900ms.

Virtual Smart Card Defined

HPEC 2012 - 23AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.

The keyd Daemon: A Virtual Smartcard

Keyd

HPEC 2012 - 24AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.– Has access to all the keys.

The keyd Daemon: A Virtual Smartcard

Keyd

HPEC 2012 - 25AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.– Has access to all the keys.

• We then created a library that conformed to the PKCS#11 standard and could talk to this daemon.– Loaded by applications running as a HPC user.

The keyd Daemon: A Virtual Smartcard

Keyd

PKCS#11

HPEC 2012 - 26AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.– Has access to all the keys.

• We then created a library that conformed to the PKCS#11 standard and could talk to this daemon.– Loaded by applications running as a HPC user.– Connects through a unix socket.– User credentials passed through the socket

Secure, provided you trust your linux kernel.

The keyd Daemon: A Virtual Smartcard

Keyd

PKCS#11

HPEC 2012 - 27AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.– Has access to all the keys.

• We then created a library that conformed to the PKCS#11 standard and could talk to this daemon.– Loaded by applications running as a HPC user.– Connects through a unix socket.– User credentials passed through the socket

Secure, provided you trust your linux kernel.

• The SVN client can then load the PKCS#11 library and use the keys to authenticate to the SVN server.

The keyd Daemon: A Virtual Smartcard

Keyd

PKCS#11

HPEC 2012 - 28AJP 9/12/2012

• We created the keyd daemon to be the brains of our virtual smartcard.– Runs as it’s own user account.– Has access to all the keys.

• We then created a library that conformed to the PKCS#11 standard and could talk to this daemon.– Loaded by applications running as a HPC user.– Connects through a unix socket.– User credentials passed through the socket

Secure, provided you trust your linux kernel.

• The SVN client can then load the PKCS#11 library and use the keys to authenticate to the SVN server.– Other applications can be enabled in the

future.

The keyd Daemon: A Virtual Smartcard

Keyd

PKCS#11

HPEC 2012 - 29AJP 9/12/2012

• The SVN server was configured to accept the LLGrid’s root of trust.

• The SVN client on the LLGrid was configured to load the keyd daemon PKCS#11 library.– One configuration entry: ssl-pkcs11-provider=libkeyd_pkcs11

Configuring SVN for TLS Client Auth

SVN User SVN Server

Connection Request

Authentication Request

Signed Authentication Responseand copy of PKI certificate

A

A

A

Keyd Daemon

HPEC 2012 - 30AJP 9/12/2012

• What is the LLGrid

• The Problem: External services authentication

• The Solution: Cryptographic authentication

• Results

Outline

HPEC 2012 - 31AJP 9/12/2012

• Keypair generation and X509 PKI certificate creation is performed during user account creation.– LLGrid Adminstrators act as the root of trust.

• We developed scripts that execute parallel key generation across nodes in the cluster.

X509 PKI Certificate Enrollment

1 10 100 10000

50

100

150

200

250

300

350

400

450

500

Serial

Parallel

Nodes

Tim

e (s

econ

ds)

Keypair & Certificate Generation

– Each certificate asserts both the user identity and the node identity to meet the guidelines to be used for either server or client TLS authentication.

HPEC 2012 - 32AJP 9/12/2012

• Created a general purpose key storage and certificate management solution for HPC.– Keys are not managed by the end-user, ensuring a low risk of

compromise requiring revocation.

• Demonstrated that it can be used to enable single sign-on integration to systems outside of the HPC.– Mitigated security concerns over passwords being stored on the

LLGrid central storage.– Avoided the issue of periodic password changes impacting batch

processing.

Results

HPEC 2012 - 33AJP 9/12/2012

• Future work will look to use these PKI certificates to secure inter-node web services communication.– Certificates are valid for both TLS client or server authentication.

Future Work

HPEC 2012 - 34AJP 9/12/2012

Questions?