Hi all,
After our catch up, we were discussing performance matters. I decided to start on this
while waiting for some of my tickets to be reviewed and to see what's going on.
These tests were carried out on a virtual machine configured in search 6 to have access to
6 CPU's, and search 12 with 12 CPU. Both machines had access to 8GB of ram.
The hardware is an i7 2.2GHz with 6 cores (12 threads) and 32GB of ram, with NVME storage
provided.
The rows are the VM CPU's available, and the columns are the number of threads in
nsslapd-threadnumber. No other variables were changed. The database has 6000 users and
4000 groups. The instance was restarted before each test. The search was a randomised uid
equality test with a single result. I provided the thread 6 and 12 columns to try to match
the VM and host specs rather than just the traditional base 2 sequence we see.
I've attached a screen shot of the results, but I have some initial thoughts to
provide on this. What's interesting is our initial 1 thread performance and how
steeply it ramps up towards 4 thread. This in mind it's not a linear increase. Per
thread on s6 we go from ~3800 to ~2500 ops per second, and a similar ratio exists in s12.
What is stark is that after t4 we immediately see a per thread *decline* despite the
greater amount of available computer resources. This indicates that it is poor locking and
thread coordination causing a rapid decline in performance. This was true on both s6 and
s12. The decline intesifies rapidly once we exceed the CPU avail on the host (s6 between
t6 to t12), but still declines even when we do have the hardware threads available in
s12.
I will perform some testing between t1 and t6 versions to see if I can isolate which
functions are having a growth in time consumption.
For now an early recommendation is that we alter our default CPU auto-tuning. Currently we
use a curve which starts at 16 threads from 1 to 4 cores, and then tapering down to 512
cores to 512 threads - however in almost all of these autotuned threads we have threads
greater than our core count. This from this graph would indicate that this decision only
hurts our performance rather than improving it. I suggest we change our thread autotuning
to be 1 to 1 ratio of threads to cores to prevent over contention on lock resources.
Thanks, more to come once I setup this profiling on a real machine so I can generate
flamegraphs.
—
Sincerely,
William Brown
Senior Software Engineer, 389 Directory Server
SUSE Labs