June 5, 2003 2:54 PM PDT
Intel: Linux good for 32-chip server
- Related Stories
HP puts Windows server back on topMay 21, 2003
Intel has been testing how well Linux can handle large databases, a demanding task that requires a server to perform huge numbers of transactions. In laboratory tests, the company found that a Linux-based 32-processor Itanium server is nearing the ability to perform 600,000 transactions per minute, a score that puts it near today's most powerful Windows and Unix servers.
Sunil Saxena, Intel's principal engineer for the Linux operating system, described the result at the Enterprise Linux Forum here Thursday. "We're very close" to reaching the 600,000 transaction-per-minute level on a database server speed test called TPC-C.
The current top score on the Transaction Processing Performance Council's TPC-C is 707,000 transactions per minute, held by a Helwett-Packard Windows server using the forthcoming Itanium 2 6M processor.
Linux currently is most widely used on lower-end servers, typically with one, two or four processors; but many databases used at the heart of corporate computing operations require more powerful machines with dozens of processors. Endowing Linux with this ability could expand Linux's market and allow it to encroach further onto the turf of Windows and Unix.
Intel began its testing on a four-processor Hewlett-Packard rx5670 Itanium server with a forthcoming version of Oracle's database software and Red Hat Enterprise Linux AS 2.1. That system produced a score of 81,000 on the TPC-C test last November, a result Oracle and others loudly trumpeted.
One holy grail of server designers is "linear scaling," under which increasing the processor count increases performance by the same amount. With Linux on the Itanium systems, Intel got close: Moving from four to 32 processors increased the processor count by a factor of eight and performance nearly as much, from 81,000 to near 600,000, Saxena said.
"It's roughly linear," he said.
Saxena was mum about many details of the 32-processor test, including whether it used the current 2.4 version of the Linux kernel--the open-source code at the heart of Linux--or the 2.5 development version that eventually will be released as 2.6 in products. He also declined to say whose server Intel used.
He did say, though, that the test involved little modification to the kernel.
When Intel and its allies conducted the four-processor test, they increased the performance from a starting score of 28,000 to 81,000 in six months, Saxena said. About 40 percent of the upgrade came through improvements in "compilation," the process by which Linux source code is translated into instructions the computer understands. The team used Intel's compiler for the test, not the GCC compiler most frequently used to compile Linux software.
Another 20 percent to 30 percent of the improvement came from changes to Linux itself, Saxena said.