Hello!
On Wed, May 25, 2016 at 10:09 PM, Shivakumar Gokaram wrote:
> Thank you for the response. I agree that running the benchmark is the best.
> I just wanted to get a feel on what is the #'s to expect in a broad range
> i.e is this 1K TPS or 100K TPS something like that.
The whole range is possible depending on your usage, data set and
traffic patterns. When you do anything wrong, we can help you optimize
things. But we cannot help with empty talks like this :)
> I am a bit surprised that get is not lock free, i am quite certain
> discussing with agentz on this and confirming it. This is interesting.
The locking effect of shdict is often exaggerated. Because it's
completely memory based and the O(log n) time complexity largely
depends on the actual data set. For most setups, you cannot even see
shdict operations in the online flame graphs. Again, it's highly
depending on your workload, dataset, and usage.
> Do we any bench mark test reports on a instance for a use case returning
> back some simple response from LUA.
>
You can do whatever benchmarks as you like. It's quicker and easier
than our lengthy discussions here. But please ensure your benchmark is
as realistic as possible since micro-benchmarks often do more harm
than good, and can mislead you to completely wrong conclusions. Just a
warning :)
Best regards,
-agentzh