Information Technology Reference
In-Depth Information
downloading a file within your own country may reach a bandwidth of multiple MBps
and a latency of less than 10ms; for foreign countries, depending on distance and
connectivity, this may decrease to kBps and latency of a few hundred milliseconds.
For a Gigabyte file this can make the difference between minutes and hours.
Latencies of milliseconds sound comparatively little considering the acceptable
delays of about 1 second for loading and displaying a website without interrupting the
user flow [4]. However, this latency is a constant, adding to any transaction between
server and client. Implicitly, whilst it has hardly any impact on large, it creates
massive delays for any interaction that bases on multiple communication exchanges
with comparatively short messages. In particular for real-time interactive applications,
such as MMO games, this built up in latency leads to massive problems.
Table 1.
Network performance in different environments
Latency
Bandwidth
Internet
150 ms
10 Mbps
Server Farms
10 ms
1 Gbs
High Performance Computing
<1 ms
up to 100 Gbs
Notably, latency and bandwidth are subject to physical constraints (cables, speed
of light, routing etc.) that start to catch up with the requirements. It is of interest to
compare the situation for average internet access with what is currently possible on
the high end of the scale. Table 1 summarises the typical network performance
parameters in three typical domains - obviously the figures may differ according to
the server's capabilities, its load etc.
What is more important than the concrete figures is however that these domains
differ by 1-2 orders of magnitude in performance, i.e. internet is 10 times “slower” (to
react) than intranet, which in turn is (more than) 10 times slower than high
performance computing. Similarly for the bandwidth, where internet's capabilities are
1/100s of the intranet, which in turn is roughly 1/100s of High Performance
Computing (HPC). We must in this context distinguish between the cloud resources
(typically server farms), and connection with the end-user. So that cloud access to the
cloud, and potentially between clouds is internet-based, whereas resources within the
cloud can reach an interconnect of the speed of server farms.
Notably cloud systems try to reduce the communication limitations further by
compensating for distance and concurrent usage through replication and relocation of
the service instances. Even so, the inherent (physical) restrictions cannot be
overcome, but are the core constraining factors. What is more, the level of freedom,
i.e. the degree of influence a management system can take, is higher with “higher-
level” domains, than with low-level ones. This is partially due to the fact that the
performance is actually achieved by maximum alignment of the system layout with
the application cases - factors such as uncontrolled concurrency, or even shared
resources are thus not supported in these domains in the first instance.
2.2
Execution System / Middleware
In order to allow for the dynamic distribution and instantiation of services,
respectively applications, it is necessary to be able to package them and re-instantiate
