Efficiency of associative array

The simulator uses various algorithms and optimisations for these things, there isn't any end-user control over them. I would suggest that you file a support ticket and share a testcase with the support engineer, so that the R&D team can investigate whether your particular case could be optimised either in the SV code or in the tool itself.

I think I've found the cause. It's really bad practice to call 'find with...' function for associative array. A benchmark of locator method for various vectors, like dynamic array, associative array and queue, tells that the performance would drop tremendously if the 'find with...' method was applied on associative array. Admittedly, if we locate the associative array through its index directly, it behaves as expected.

Though it's not likely to locate the associative array through 'find with...' method instead of its intuitive index, there's chance when users have to apply some complex expression to 'find with...' method, e.g, a_array.find with (item.addr >= MIN; item.addr <= MAX; item.attr == 1;...), where as a result the performance might heavily be impacted. In this case, it's embarrassed when user initially takes advantage of associative array for efficient locating and some time later, he/she somehow has to apply 'find with...' method on the associative array as well.

Do you have any suggestion on how to balance the pros and cons of associative array?

Seems like you're looking for a free lunch. 'find with' is inherently expensive, because it has to loop over all elements to find which ones satisfy your expression.

For your 'a_array.find with (item.address == end_addr)' example that is slow you're comparing apples against oranges. On one side you have an array indexed by address and on the other side you have an array indexed by the item reference itself. You can't draw any conclusions regarding performance from such an experiment...

The point is that 'find with' is slow especially for associative array.
If apply 'find with' on queue, dynamic array, and associative array respectively with the same expression, we can see the queue and dynamic array cost almost the same time while the associative array spent much more.

The point is that 'find with' is slow especially for associative array.
If apply 'find with' on queue, dynamic array, and associative array respectively with the same expression, we can see the queue and dynamic array cost almost the same time while the associative array spent much more.

Set is inherently less iteration friendly than list.

Thank you for you comments.

Incisive is a legacy product and isn't being developed any more. Its replacement, Xcelium, has many enhancements and optimisations, so if you are concerned about performance, please consider upgrading to Xcelium.

I would also suggest that rather than worrying about benchmarking individual language features, you'd be better off profiling your real testbench and seeing what are the biggest time consumers overall for a full regression suite. You will probably find that the associative array look-ups are not significant compared to other code that you've written, that you could optimise yourself. Even if VCS is 3x faster to index an associative array, if it's only taking maybe 1% of the simulation time (random number in lieu of real data) then making it 3x faster will save you 0.66% of your regression time. Hardly worth shouting about

Finally, please, for your own benefit, think carefully before posting comparative benchmarking data in public, you might well be violating licence conditions from Cadence and Synopsys, which could land you or your employer in trouble...

Thank you for your kind suggestion and sorry for my haste. I will recall my post for the comparative data.

Yes, you're right. The profiling shows I've called so much uvm_top.find() in my testbench, which is the bottleneck for my regression time.