Server computers for web applications are typically configured from
systems based from commodity processors. To gain sufficient power, these
processors are in systems connected by shared memory (SMP) interconnects
and/or aggregated into cluster computers connected by fast Ethernet or
In this talk, I will explain why, from the point of view of
web servers (and for that matter any transaction-oriented application),
this is not only inefficient in
terms of up-front costs but ongoing energy costs. This arises for a
number of reasons: firstly the processors have a high clock speed, which
causes high power consumption, but gives little compensating benefit for
transactional applications. Secondly, much of the processor's chip area
(and power consumption) is oriented towards extracting a high-degree
instruction-level parallelism, which cannot be utilized for web applications.
Thirdly, these systems waste energy because they are comprised
of many disks connected by interconnects. Finally, in most instances,
a single physical server system is devoted to a single or small numbers of
applications for an organization, even when the average demand for those
services is quite low.
I will then explain why the new wave of aggressive multicore /
multithreaded processors are, such as the T2 CoolThreads, are a much more
cost-effective solution for transaction-oriented applications. This is
because they are specifically designed to exploit the concurrency of
handling a large number of simultaneous transactions, while not wasting
chip area (and energy) on architectural features that bring no benefit
for these applications. By providing this processing power in a single
chip, a powerful server can be contained in a single rack unit, with a
relatively small number of disks and modest interconnection requirements.
When combined with the technology of virtualization, it is also
possible to consolidate different (logical) server systems on the
same physical server. With state-of-the-art virtualization providing
high degrees of security, such a configuration can
provides an efficient and flexible solution for meeting peak demand.