642
M. Delfino
raw data. Developments in the LHC domain, such as the LHCb turbo stream, 93 may
also reduce the need for large-scale wide-area distributed processing. In parallel,
as discussed below, data centers, particularly commercial ones, are growing in
capacity. It is quite conceivable that the future needs of many experiments could
be serviced by just a few data centers.
Another issue is the foreseen disappearance of magnetic tape technology. Tape
has already disappeared from the commodity market, driven by ever lower disk
prices and alternative backup media such as DVDs. If electrical energy considerations are excluded, the overall cost of high-end tape, associated robotic mechanisms
and periodic migration to new media is becoming less attractive. Nevertheless, as
of this writing, tape storage continues to be more economical than disk storage,
especially if electrical costs are taken into account. The prospects are worrisome,
however, as fewer companies continue to do R&D in tape technology. Hence, the
particle physics community should be preparing for a hypothetical tape-less future
by adapting its data management and cluster architectures as described above,
allowing for multiple disk copies of the same dataset on a cluster or across a Grid
or in several Clouds.
Grid and Cloud computing have brought along a resurgence of “timesharing”, 94
with research data centers deploying large clusters which give services to many
projects of which only a minority are particle or astroparticle physics projects. The
early adoption of Grids by particle physics, however, has resulted in development
and deployment of a number of cluster and Grid tools which require extremely specific, non-commercially supported operating system or data server configurations
which are of no interest to other projects or to the managers of general purpose
clusters or Clouds. This does not mean that the developments led by particle physics
are not worthwhile or of good quality; it simply means that further steps must be
taken outside the particle physics scope to ensure a sustainable future and general
applicability. Therefore, the particle physics community has to share knowledge
about its developments, but also continuously evaluate alternative solutions. A
case in point may be the use of non-standard data serving packages by LHC
sites: it is perceived as the best solution (especially for LHC Tier-1 centers), but
smaller particle or astroparticle projects with much lower data volumes often prefer
what they perceive as a simpler, more standard solution such as NFS. Standard
deployment of NFS on a cluster with thousands of CPU cores will however not
be favored by system managers or commercial providers for a number of practical
operational reasons, and they will propose newer technologies such as Ceph 95
93 Benson, Sean, et al. “The LHCb turbo stream.” Journal of Physics: Conference Series. Vol. 664.
No. 8. IOP Publishing, 2015.
94 The term “timesharing” was coined in the 1960s to describe the simultaneous use of a single
“mainframe” computer for multiple batch or interactive tasks by multiplexing processes into the
mainframe CPU.
95 Weil, Sage A., et al. “Ceph: A scalable, high-performance distributed file system.” Proceedings
of the 7th symposium on Operating systems design and implementation. USENIX Association,
2006.
M. Delfino
raw data. Developments in the LHC domain, such as the LHCb turbo stream, 93 may
also reduce the need for large-scale wide-area distributed processing. In parallel,
as discussed below, data centers, particularly commercial ones, are growing in
capacity. It is quite conceivable that the future needs of many experiments could
be serviced by just a few data centers.
Another issue is the foreseen disappearance of magnetic tape technology. Tape
has already disappeared from the commodity market, driven by ever lower disk
prices and alternative backup media such as DVDs. If electrical energy considerations are excluded, the overall cost of high-end tape, associated robotic mechanisms
and periodic migration to new media is becoming less attractive. Nevertheless, as
of this writing, tape storage continues to be more economical than disk storage,
especially if electrical costs are taken into account. The prospects are worrisome,
however, as fewer companies continue to do R&D in tape technology. Hence, the
particle physics community should be preparing for a hypothetical tape-less future
by adapting its data management and cluster architectures as described above,
allowing for multiple disk copies of the same dataset on a cluster or across a Grid
or in several Clouds.
Grid and Cloud computing have brought along a resurgence of “timesharing”, 94
with research data centers deploying large clusters which give services to many
projects of which only a minority are particle or astroparticle physics projects. The
early adoption of Grids by particle physics, however, has resulted in development
and deployment of a number of cluster and Grid tools which require extremely specific, non-commercially supported operating system or data server configurations
which are of no interest to other projects or to the managers of general purpose
clusters or Clouds. This does not mean that the developments led by particle physics
are not worthwhile or of good quality; it simply means that further steps must be
taken outside the particle physics scope to ensure a sustainable future and general
applicability. Therefore, the particle physics community has to share knowledge
about its developments, but also continuously evaluate alternative solutions. A
case in point may be the use of non-standard data serving packages by LHC
sites: it is perceived as the best solution (especially for LHC Tier-1 centers), but
smaller particle or astroparticle projects with much lower data volumes often prefer
what they perceive as a simpler, more standard solution such as NFS. Standard
deployment of NFS on a cluster with thousands of CPU cores will however not
be favored by system managers or commercial providers for a number of practical
operational reasons, and they will propose newer technologies such as Ceph 95
93 Benson, Sean, et al. “The LHCb turbo stream.” Journal of Physics: Conference Series. Vol. 664.
No. 8. IOP Publishing, 2015.
94 The term “timesharing” was coined in the 1960s to describe the simultaneous use of a single
“mainframe” computer for multiple batch or interactive tasks by multiplexing processes into the
mainframe CPU.
95 Weil, Sage A., et al. “Ceph: A scalable, high-performance distributed file system.” Proceedings
of the 7th symposium on Operating systems design and implementation. USENIX Association,
2006.