Data Management Expert Panel

fullgorgedcutΔίκτυα και Επικοινωνίες

24 Οκτ 2013 (πριν από 4 χρόνια και 15 μέρες)

155 εμφανίσεις

Data Management Expert Panel

RLS

Globus
-
EDG Replica Location Service


Joint Design in the form of the ‘Giggle’ architecture


Reference Implementation by Globus Team within GT2


Focus on performance and features


Implementation by EDG team in a Web Services Framework


Focus on manageability and robustness


No interoperability due to differences in communication protocols
and language bindings


EDG implementation chosen to build grid catalog for POOL


January 2003



RLS then (Jan 2003) : EDG & Globus Impl.


Globus RLS


C
-
based daemon
-
style technology


C
-
language binding, java through
JNI



MySQL
-
only backend implementation



Supports LRCs and RLIs.





Uses proprietary Globus Toolkit 2
(GT2) protocols for network
communications



N:M logical to physical filename
mapping



Schema is not designed to support
GUIDs and aliasing



Evolution is very hard due to
hardcoded schema and SQL, code
change required


WP2 RLS


Java
-
based technology


Native C, C++, Java, Perl, Python

bindings


MySQL and Oracle support, easy to
extend to more DBMS


LRC only at the moment. Support
planned for RLIs.


Uses Web Service protocols for
network communication Small client,
no dependencies (on GT2 or others)


N:1:M logical to physical file
mapping


Schema has natural support for
GUIDs and alias
-
aliasing



Evolution is easier, no code change
necessary. and SQL, SQL is in
configuration file.

RLS now (June 2002) : EDG & Globus Impl.


Globus RLS


C
-
based daemon
-
style technology


Native C and java bindings




MySQL and Postgres backend
implementation




Supports LRCs and RLIs.



Uses proprietary Globus Toolkit 2
(GT2) protocols for network
communications



N:M logical to physical filename
mapping



Schema is not designed to support
GUIDs and aliasing



Evolution is very hard due to
hardcoded schema and SQL, code
change required


WP2 RLS


Java
-
based technology


Native C, C++, Java, Perl, Python

bindings


MySQL and Oracle support, easy to
extend to more DBMS


Supports LRCs and RLIs.


Uses Web Service protocols for
network communication Small client,
no dependencies (on GT2 or others)


N:1:M logical to physical file
mapping


Schema has natural support for
GUIDs and alias
-
aliasing



Evolution is easier, no code change
necessary. and SQL, SQL is in
configuration file.

RLS : Which one for which user?


Differences between functionality growing smaller


Commitment on both sides to implement new functionality in a
interoperable way e.g. bulk upload, new query mechanisms


WP2 / CERN IT
-
DB not able to support external (non
-
EDG / LCG)
customers


Deployment model still different


Major outstanding technical difference, which will be resolved with GT3


Choice probably comes down to what components you already use

Interoperability


Had meeting with Globus after CHEP


Agreed to do it now would require lots of extra work


wrapper code to
hide differences in network protocols


GT3, and GGF standards will make this easier


Second meeting scheduled mid July 2003


Interoperability a strategic goal


Aim for full interoperability within the context of OGSA


WP2 Work Schedule

EDG WP2 Work Schedule (1/3)


April 2003 : EDG 2.0
-

First release of new data management
framework


Services Deployed


Replica Location Service (Local Replica Catalog)


Replica Metadata Catalog


Replica Optimisation Service


Replica Manager


Issues:


No security integration (authentication + authorization)


Single Local Replica Catalog/ Replica Metadata Catalog per VO



EDG WP2 Work Schedule (2/3)


July 2003 : EDG 2.1


Focus on missing functionality from EDG 2.0


July 11


Replica Location Indices


LRC pushes updates to registered RLIs


EDG Replica Manager supports multiple LRCs and RLIs


July 22


Security


Integrate VOMS


Deployment of EDG Trust Manager into tomcat


authentication
available for all java based services. R
-
GMA, SE also using this.


Deployment of EDG Authorization Manager to allow services to make
course grained authorization decisions



EDG WP2 Work Schedule (3/3)


October 2003


RLS Service Proxy
-

Hide the interaction with


RLI and remote LRCs


Information service


Removes complexity and duplicated code from


EDG Replica Manager


POOL


Grid File Access Library


The Grid look like two “Local” Replica Catalog


One is LRC for local site


One acts as a proxy for all other LRCs in the grid





Problems in current architecture (1/2)


RLS complexity


The RLS will have a large set of LRCs, RLIs running across many sites


Currently the client (e.g. EDG Replica Manager or POOL) will need to
manage these interactions


Client failures


All failures are managed at the client side


if the client itself fails, there is no means of recovery
-

no state can be
read back


Scalability of transfer


If each client is allowed to issue GridFTP requests in a fabric, the
network will be saturated


Currently no way to find out whether a given file is already being
replicated

Future Work

Problems in current architecture (2/2)


Outbound connectivity


The worker nodes need to have direct outbound connectivity for the
Replica Manager to work.


This is not a given in all fabrics


Failure upon unreachable remote service (RMC)


The RMC will be deployed only at one site



This means jobs can fail if network between site and RMC breaks


Possible Solutions (1/2)


RLS complexity


Service for RLS which acts as a proxy to RLI and all remote LRCs


Client failures


Better client side libraries which handle network retries


Hide/manage the network and service related exceptions


Scalability of transfer


Single service at each site could schedule replications, and block
requests for that file until it arrives



Possible Solutions (2/2)


Outbound connectivity


SOAP proxy solves problem for all services


Data transfer (gridftp) still a problem


Failure upon unreachable remove service (RMC)


WAN
-
Distributed database


Distributed messaging system to store actions at worker node site and
handle retries


Use vendor supplied solutions rather than re
-
invent


Outstanding Components


Collection management


Confined Collections


Limited to PFNs all stored on the same LRC


Could be considered as ‘directories’


Allows a user to replicate a set of PFNs from one site to another


Free Collections


PFNs which could be stored on any site


Are these needed? Use cases not clear


Replica Subscription Service


Provides functionality of GDMP from EDG Release 1


Allows background third party replication between SRMs

GFAL

Grid File Access Library


A solution to the “Grid Open” problem


POSIX API for opening a file in the Grid


Hides the complexity of


Replica Metadata Catalog


Replica Location Service


Storage Resource Manager


MSS backends


File access Protocols


file, rfio, dcap, root,


GFAL Overview


Physics Application

Replica

Catalog

Client

SRM
Client

Local

File
I/O

rfio
I/O

dCap
I/O

Grid File Access Library (GFAL)

SRM

Service

dCap

Service

rfio

Service

RC

Services

MSS

Service

Local

Disk

POSIX I/O

Wide Area

Access


VFS

root
I/O

GFAL example (open)