xmallox-test benchmark 40x slower with tcmalloc than with glibc's malloc

[JulianMarcusSchroeder]

This is a corner case that may be of interest.
I run on a 32 core AARCH64 Neoverse v1 system / Ubuntu-22 with linux-5.15/linux-6.2
The xmalloc-test runs 32 malloc and 32 free threads that concurrently allocate/free 409600x64 bytes.
Repro:
git clone https://github.com/daanx/mimalloc-bench
cd mimalloc-bench
sed -i "s/if defined(APPLE)/if 1/g" bench/alloc-test/test_common.h
./build-bench-env.sh tc bench
cd out/bench
SYSMALLOC=1 ./xmalloc-test -w 32 -t 5 -s 64
rtime: 1.085, free/sec: 92.164 M
LD_PRELOAD=/home/ubuntu/gperftools/.libs/libtcmalloc_minimal.so.4 ./xmalloc-test -w 32 -t 5 -s 64
rtime: 39.578, free/sec: 2.527 M
This is with gperftools-2.13.

[alk]

Thanks for letting us know. So I can argue that this specific benchmark is somewhat artificial in that it has threads releasing bunches of objects into same glibc malloc heap "partition". Our design is non-partitioned heap which actually works surprisingly well across lots of workloads, but occasionally loses quite a bit as highlighted in this workload.

But there are definitely lessons for us too. Thanks again.