Comment # 39 on bug 1192457 from Martin Wilck

Quick test on 5.14.14, x86_64:

The last column is the size of the modules tree after the command
in the first column. The other columns indicate times and max RSS.
446MiB is the uncompressed size. All tests run consecutively, i.e. from the
page cache.

command                         real    sys     usr     mem     size
"xz"                                                            143
"xz -d"                         0:07.58 0.56    7.01    4268    446
"xz -0"                         0:20.15 0.49    19.65   4880    151
"xz -d"                         0:08.05 0.55    7.49    2424    446
"zstd -21 --rm -q -T0"          2:21.79 0.95    140.75  114156  153
"zstd -d --rm -q"               0:01.71 0.60    0.89    11216   447
"zstd -15 --rm -q -T0"          0:49.35 0.77    48.53   96104   158
"zstd -d --rm -q"               0:01.33 0.65    0.67    7488    447
"zstd -10 --rm -q -T0"          0:11.17 0.54    10.59   40760   162
"zstd -d --rm -q"               0:01.24 0.54    0.66    7436    447
"zstd -6 --rm -q -T0"           0:05.79 0.53    5.30    32388   166
"zstd -d --rm -q"               0:01.28 0.49    0.77    5336    447
"zstd -3 --rm -q -T0"           0:03.06 0.51    2.56    25324   171
"zstd -d --rm -q"               0:01.88 0.79    1.04    5460    447
"zstd -0 --rm -q -T0"           0:03.24 0.58    2.69    25404   171
"zstd -d --rm -q"               0:01.24 0.52    0.71    5460    447

xz -6 (standard level) compression takes about 2:20 minutes. 
You can see that decompression time is actually constantly low with zstd.

zstd -10 wins about 9MB, or 6% relative to the default size, compared to zstd
-3, taking 4x as much time. Given that the compression time matters only during
kernel building, and "zstd -10" is still 14x faster than "xz -6", it might be a
good compromise. We could actually go for "zstd -21 -T0": The compression time
is about the same as xz -6, and we'd loose just ~6% disk space.