Playback performance depends on three bottlenecks (in order of significance): network quality, client load, and server load. In this section we talk about each of these and how to determine which bottleneck requires attention.
Network quality can be measured as a combination of latency, throughput, and stability. When determining network quality you may want to run Scyld Cloud Workstation on its own to guarantee that other applications or clients are not consuming large amounts of network resources at the same time.
Latency between the client and server can be measured using
times. Acceptable latency depends on the applications being used.
CAD users, for example, may find ping times up to 150 ms to be quite
usable and 300 ms to be usable for sporadic use. Testing and
demoing of applications like Google Earth are typically over 802.11g
connections with ping times of 30-80 ms.
When running fullscreen animations at 1440x900, Scyld Cloud Workstation has a typical throughput consumption of 4 Mbps. Throughput consumption drops dramatically when pixels on the screen do not change. We conservatively recommend 5.5 Mbps. This is typically not a bottleneck for Scyld Cloud Workstation since it’s common for clients and servers to have more than 4 Mbps of bandwidth, but it is still worth remembering.
Decoding is largely dependent on the web browser implementation and the CPU performance of the client. We recommend using Chrome as it performs best with Scyld Cloud Workstation in testing.
CPU performance depends on the hardware and the load on the system. We test on modern CPUs such as the multi-core Intel i5s and i7s from 2011 and later. When evaluating playback performance, verify that other applications are not also consuming large amounts of CPU time.
Decreasing screen resolution on the server-side is another option for reducing load on the client. While we recommend 1600x900, users may find that 1280x720 offers a better overall experience.
If you are running the non-WebGL version of Scyld Cloud Workstation, performance is expected to be considerably slower (depending on the CPU). Lowering the remote server’s screen resolution and using Chrome is strongly recommended in this case.
Server load is typically not a large bottleneck since Scyld Cloud Workstation does not consume much server-side CPU time. GPU consumption does increase, but for NVIDIA GRID cards the display capture and encoding is done on a part of the GPU that is independent of computation.