Hardware Reference
In-Depth Information
ensures that compute, storage, and networking are balanced across the Cen-
ter. Achieving balance means they are continuously augmenting both capacity
and capability of the centralized storage systems based on improvements to
computational and network capabilities.
NERSC is able to provide both external and internal users and collabo-
rators with high-bandwidth access to data because of their security stance.
In general, NERSC's network infrastructure eliminates up-front barriers when
users or collaborators want to gain access to NERSC data and computational
systems that are guided by each science team's desires. Instead of using re-
walls as their main security policy, which impedes all data transfer perfor-
mance, NERSC sponsors a unique network analysis-based security program
called Bro [1] that actively monitors and enables or disables real-time access
to systems and information at the facility.
The Center's computing resources are currently deployed at a multi-
petaflop scale. The two largest systems, Hopper and Edison, together have
over 250,000 cores, 550 TB of memory, and 3.5 PFLOPS of peak computa-
tional capability. These systems offer users the highest concurrency and are
used at full scale.
NERSC has several data-centric or mid-range computing systems that
continue to serve user's needs. The primary data-centric systems are PDSF
and Genepool. They serve the high-energy and nuclear physics community,
and the Joint Genome Institute,
2
respectively. The primary mid-range system
is Carver which serves the broader NERSC user community. Figure 2.1 (JGI)
shows all the computing and storage resources available at NERSC.
The compute systems are regularly utilized at 98% of their capability with
a variety of jobs from all areas of science executing simultaneously on a node
or on a system of nodes, depending on the particular compute system queue
configuration. In conclusion, there is no particular workload that the compute
or storage systems predominantly serve.
2.2 I/O Hardware
NERSC has focused on providing storage systems and services to cover
a variety of I/O needs, from requirements of aggressive I/O to sharing data
with collaborators around the world. Local scratch file systems are deployed
as part of a particular computational system, and thus have about a five-year
lifespan. The global or centerwide file systems that are part of what we call
the NERSC Global Filesystems (NGF) are managed by an evergreen strategy
where existing hardware is augmented and refreshed on a regular basis, so as
2
JGI is a DOE facility focused on advanced genetic research operated by the University
of California.
