Build times for our design

[oharboe]

There is no question here, just a demonstration of the use of the build time summary for a design that has one top level macro and 5 macros at the top level. What is being built here are mock abstracts for the macros at the top level as we are only interested in regression testing the top level on our build servers.

The top level is the slowest in detailed routing and we have 5 macros at the top level.

We are not currently building the macros at the top level in parallel, though we intend to, but it would only cut the running time by 50%.

This is the command I run to build a summary of all the build times find -regex ./logs/.*/base -type d -exec python3 util/genElapsedTime.py -d {} \; | sort -hr -k2

Stage	time/s
5_2_TritonRoute	21313
5_2_TritonRoute	7284
5_2_TritonRoute	4744
5_2_TritonRoute	2964
5_2_TritonRoute	1825
5_2_TritonRoute	1482
3_3_place_gp	1274
4_1_cts	1158
3_3_place_gp	666
1_1_yosys	617
4_1_cts	379
6_report	332
3_3_place_gp	312
5_2_TritonRoute	294
3_5_opendp	293
1_1_yosys	287
5_2_TritonRoute	282
3_3_place_gp	278
4_1_cts	273
3_4_resizer	224
4_1_cts	203
5_1_fastroute	198
6_report	164
5_1_fastroute	164
1_1_yosys	160
5_1_fastroute	142
3_1_place_gp_skip_io	132
5_1_fastroute	126
3_4_resizer	123
6_report	116
3_5_opendp	109
...	...

[maliberty]

That is average behaviour as drt is costly. It does scale pretty well with threads so more CPUs helps. We do have some work to do distributed drt if you want to test it out but it isn't in the flow currently due to the extra setup overhead.

[maliberty]

How many threads in your data above?

[oharboe]

12, it's on a local somewhat dated workstation, not on our newer servers.

Ref my earlier plot, it looked like there's little gain beyond 32 threads: #3298

[oharboe]

No question or suggestions here, just a datapoint that I thought it might have some value in sharing.

[maliberty]

Fwiw we test with 40 cpus and I know Google often uses 96. We have tested with several hundred in distributed mode.

[oharboe]

We use 128 threads, but at 32, the gains taper off.

Is this not what you are observing?

[maliberty]

It is somewhat dependent on design size. Larger designs should see more benefit.

[oharboe]

👍

[QuantamHD]

Another important factor to mention is that early in DRT you leverage a lot of parallelism, but towards the end you're left with just a few violations that can often be in single workers.