performance.txt

# c with no concurrency

$ time ./pbf 
total node references 93796649
real	0m17.815s
user	0m16.597s
sys	0m1.208s

# python with no concurrency

$ time ../street.py
done. total node refs: 93796649
real	2m10.270s
user	2m57.327s
sys	0m8.181s

(7.2x speedup for C)

In Python version using imposm.parser.OSMParser concurrency does not seem to help.
IDF:
  concurrency 4 25.5s
  concurrency 2 36.5s
Benelux:
 concurrency 2 1m56s
 concurrency 1 2m12s (~500MB resident)

# C with 8 parallel processes on Benelux feed
$ for i in {1..8}; do time ./pbf /var/otp/graphs/benelux/planet-benelux.osm.pbf& done
real 0m35s

# C with 4 parallel processes on Benelux feed
real 0m21s
Memory use around 2-5MB per process, probably mostly for zlib buffer.

# C with 2 parallel processes on Benelux feed
real 0m18.5s

So it is entirely possible to fork and count nodes, find intersections, load nodes simultaneously.
Or, each process could touch only the blocks mod its pid == 0. The problem with the latter is that you have to merge the results, whereas the former works by itself.