Most Network Data Sits Untouched
Statistically speaking, most data on enterprise networks rarely gets accessed
after it is written to network storage, according to researchers from NetApp
Inc. and the University of California (UC). Evidently, we are too busy writing
new data to go back over old data.
Andrew Leung, a computer science researcher at the UC, presented the findings
at the USENIX conference in Boston last week. Given those results, organizations
might want to consider moving much of their data to slower but less expensive
storage units since it rarely gets accessed, he said.
The team studied the traffic that flowed through NetApp's enterprise file servers,
which manage more than 22T of material relating to all aspects of the company's
Leung said the study is the first large-scale examination of network traffic
patterns. "How people have been deploying network file systems has been
changing over the past five to 10 years," he said. "They are being
used more commonly for different kinds of things. So what we would like to know
is how this affects the workloads of the network."
During the three-month period that the network was under scrutiny, more than
90 percent of the material on the servers was never accessed. The researchers
captured packets encoded using the Common Internet File System protocol, which
Microsoft Windows applications use to save data via a network. About 1.5T of
data was transferred.
"Compared to the full amount of allocated storage on the file servers,
this represents only 10 percent of data," Leung said. "[This] means
that 90 percent of the data is untouched during this three-month period."
Moreover, among the files that were opened, 65 percent were only opened once.
And most of the rest were opened five or fewer times, though about a dozen files
were open 100,000 times or more.
"What this suggests, in general, is that files are infrequently re-accessed,"
The team also observed that the ratio of data being read from storage versus
the amount of data written to storage has changed from what had been seen in
previous studies. Bytes written compared to bytes read by a ratio of about 2-1.
"Past read-write ratios saw read-to-write ratios of 4-1 or higher,"
Developers of file systems might want to take into consideration the fact that
their creations are spending almost as much time writing data as reading data.
"The workloads are becoming more write-oriented, so the decrease in read-only
traffic and the increase in write traffic suggests that file systems want to
be more write-oriented," Leung said.
File server vendors also might want to consider re-jiggering their pre-fetching
and caching algorithms to improve performance, given those findings. "If
we know that files aren't frequently re-accessed, what this suggests is that
[caching] algorithms may not be the best for network file systems" because
the material cached will probably not get retrieved, he said.
Speaking to Government Computer News after the presentation, Leung described
the 10 percent of data that was being re-accessed. Typically, it is in the file
format most closely associated with the user's job. Architects might use computer-aided
design files, while developers use source-code files. Also, files that are higher
up in a file path or closer to the user's home directory tend to be accessed
more often than those buried deeper down in a hierarchy of subfolders.
More than 75 percent of the files being opened were very small -- less than
20K each -- although another 12 percent were more than 5G each.
Joab Jackson is the chief technology editor of Government Computing News (GCN.com).