Sun Microsystems Netscape Enterprise Server, Administrator'S Manual

Page: 262 / 532

Sizing Issues
262 Netscape Enterprise Server Administrator’s Guide
Sizing Issues
This section examines subsystems of your server and makes some
recommendations for optimal performance:
• Processors
• Memory
• Drive Space
• Networking
Processors
On Solaris and Windows NT, Enterprise Server transparently takes advantage of
multiple CPUs. The effectiveness of multiple CPUs varies with the OS and with
the nature of the workload. In general, Solaris scales better than Windows NT.
The CGI, static SSL, and mixed SSL workloads scale dramatically better on
Solaris than on NT, while NSAPI, static HTML, and the mixed workload scales
within a few percent better or worse on Solaris than on Windows NT.
With static HTML, Windows NT has some non-standard extensions to the
WinSock API that allow static files to be transmitted more efficiently than the
BSD sockets API can accomplish. These extensions also allow the system to
spend more time in the NT kernel, where SMP scaling has apparently been
better optimized than on Solaris. The mixed workload is clearly dominated by
static HTML, so it scales similarly. In the case of NSAPI, more time is spent in
the NSAPI shared library, which is functionally identical on Windows NT and
Solaris - so the scaling is similar.
The Solaris SSL test indicates that Solaris can actually exceed ideal SMP scaling -
this is probably due to improved L2 cache coherency, so that the additional 3
CPUs effectively yield an L2 cache four times as large.
In general, SSL and CGI gained the most from additional CPUs, while static
pages gained the least. This is partly due to the fact that static performance is
very fast, and couldn’t get much faster without saturating the test network. It’s
also due in part to the fact that static HTML pages are mapped to shared
memory, so cache flushes impact performance more heavily than in the more
parallelizable SSL case.