Hi, i added fast serialization to the benchmark. I checked the kryo bench source to get an idea of how much tweaking is done. The test does the same as kryo "basic-unoptimized". I preregister classes, set the serialization mode to "no shared references" and reuse my stream instances. Unsage mode is turned off (does not improve results with this bench anyway).
On my i7 fast serialization is slightly faster than kryo, but produces ~20% larger output (however in other testcases its vice versa)
I see no point adding "manual optimized" flavours, they'd be probably in the same ballpark as kryo.
I think it would be interesting to compare "real serializers" (with cycle detection an shared references) without tweaking (no pregeistering of classes, no unshared mode). In reality most users of serialization cannot or will not preregister classes .. drop me a line if interested.
Hi, i added fast serialization to the benchmark. I checked the kryo bench source to get an idea of how much tweaking is done. The test does the same as kryo "basic-unoptimized". I preregister classes, set the serialization mode to "no shared references" and reuse my stream instances. Unsage mode is turned off (does not improve results with this bench anyway). On my i7 fast serialization is slightly faster than kryo, but produces ~20% larger output (however in other testcases its vice versa) I see no point adding "manual optimized" flavours, they'd be probably in the same ballpark as kryo.
I think it would be interesting to compare "real serializers" (with cycle detection an shared references) without tweaking (no pregeistering of classes, no unshared mode). In reality most users of serialization cannot or will not preregister classes .. drop me a line if interested.
kind regards, ruediger