Database Reference
In-Depth Information
Request:
Get multiple files
from source
Source SRM
(performs reads)
Target SRM
(performs writes)
Get file
Disk
Cache
Disk
Cache
NCAR-MSS
Request to get file
Allocate space
Stage to disk
File is ready
Allocate space
Send me the file
Transfer file
Archive to MSS
Release space
Release space
Request next file
Figure 3.6
Sequence of operations between two cooperating SRMs to sup-
port robust file replication.
SRMs are also capable of performing the same operation of copying of an en-
tire directory, by first getting the directory structure from the source system
under SRM, establishing the same structure in the target system under SRM,
and then starting the transfer process into the corresponding locations in the
directories. This amounts to performing the Unix recursive remote copy com-
mand “rcp -r directory” between two remote sites for very large directories
by using file streaming.
We note that the file staging and transfer can overlap in time for multi-
ple files, and therefore there may be multiple files ready in the cache to be
transferred, creating the opportunity to transfer multiple files concurrently.
Another interesting property is that the entire process is self-regulating; that
is, if the network is too slow, the quota in the source disk gets full, and it stops
requesting files to be staged. Similarly, if the target disk gets full because the
archiving is too slow, it stops asking the source to send files. Thus, there is
no possibility for files to get lost. This setup has proven to work reliably even
in the face of long maintenance periods (usually an hour or so) of the MSSs.
In particular, this has been used for daily robust large-scale data replication
for the STAR experiment
32
between the MSS at Brookhaven National Labo-
ratory and Lawrence Berkeley National Laboratory since 2004. The volume
needed to be transferred is about 10,000 files (about 1 TB) per week. This
arrangement resulted in a 50X reduction in the error rates, from 1% to 0.02%.
The same arrangement has been used periodically by the Earth System Grid
(ESG),
33
as shown in Figure 3.6.
