Fundamentals of Quantitative Design and Analysis - Computer Architecture: A Quantitative Approach - page 55

FIGURE 1.18 Three Dell PowerEdge servers being measured and their prices as of
August 2010 . We calculated the cost of the processors by subtracting the cost of a second
processor. Similarly, we calculated the overall cost of memory by seeing what the cost of extra
memory was. Hence, the base cost of the server is adjusted by removing the estimated cost
of the default processor and memory. Chapter 5 describes how these multi-socket systems
are connected together.
Note that due to the forces of benchmarking (see Section 1.11 ), these are unusually con-
igured servers. The systems in Figure 1.18 have litle memory relative to the amount of com-
putation, and just a tiny 50 GB solid-state disk. It is inexpensive to add cores if you don't need
to add commensurate increases in memory and storage!
Rather than run statically linked C programs of SPEC CPU, SPECpower uses a more mod-
ern software stack writen in Java. It is based on SPECjbb, and it represents the server side
of business applications, with performance measured as the number transactions per second,
called ssj_ops for server side Java operations per second . It exercises not only the processor of the
server, as does SPEC CPU, but also the caches, memory system, and even the multiprocessor
interconnection system. In addition, it exercises the Java Virtual Machine (JVM), including the
JIT runtime compiler and garbage collector, as well as portions of the underlying operating
system.
As the last two rows of Figure 1.18 show, the performance and price-performance winner is
the PowerEdge R815 with four sockets and 48 cores. It hits 1.8M ssj_ops, and the ssj_ops per
dollar is highest at 145. Amazingly, the computer with the largest number of cores is the most
cost effective. In second place is the two-socket R815 with 24 cores, and the R710 with 12 cores
is in last place.
While most benchmarks (and most computer architects) care only about performance of
systems at peak load, computers rarely run at peak load. Indeed, Figure 6.2 in Chapter 6
shows the results of measuring the utilization of tens of thousands of servers over 6 months at
Google, and less than 1% operate at an average utilization of 100%. The majority have an av-
erage utilization of between 10% and 50%. Thus, the SPECpower benchmark captures power