|
|
Home |
Interfaces and Architecture |
Fault Tolerant Baseline |
High Performance |
Downloads |
High Performance Design
Statistics
Average Latency for all Invocations – 12.922 ms
Average Latency for only Faulty Invocations – 4544 ms
Minimum Jitter in the Presence of Faults – 387.59 ms
Maximum Jitter in the Presence of Faults – 7319.6 ms
Description – In this pie chart we can see the breakdown of time for our system when faults are injected. This pie chart shows the data for an average over 25 different faults. The client delay comes mostly from a built-in sleep command that is in the client to prevent constant polling of the naming service. This also includes the time for the initial attempt and exception, and any time after the replication manager finishes fixing the replica and the client is still sleeping. These breakdowns are either impossible to obtain due to clock skew or irrelevant since we cannot control fault injection time and, therefore, are not shown here.
Plans for
Phase IV
- Optimize Client Delay
- Pre-establish connections
to all replicas
- Run a thread in the
background that maintains these connections
- Tweak built-in delays in
the client
- Specifically, this will
look at the sleep command we have in the client that prevents it from
constantly polling the server and to prevent us from talking to the
naming service too much since this is a slow procedure.
- Optimize Fault Detection
Time
- Tweak heartbeat frequency
- Make the frequency as
quick as it can be without requiring too much processor time and
slowing down the CPU
- Look into forcing the
replication manager to check the primary more often than the backups for
failures so that fault detection is quicker
- Optimize Fault Correction
Time
- Tweak frequency of second
level of detection in replication manager that occurs before corrections
- This frequency is built
into a separate thread that handles re-launching a failed replica. By altering this, we can get a
quicker re-launch time.
- Also look into possibly
eliminating this delay entirely by using more interrupt-like
communication with this thread.
Latency Plots
