Memory leak P-1 about gpuowl HOT 16 CLOSED

That should happen sooner on the RX580 which has 8GB RAM.

from gpuowl.

Doing the pm1 test right now on RX580, lets see how soon it occurs.
However it did find "no factor", will post the gpuowl.log later.

from gpuowl.

Does this concern GPU memory or main system memory?

from gpuowl.

OK. Could it be that there was not enough memory on the GPU, for example because something else (or another instance) was running at the same time taking up some GPU memory?

P-1 allocates by default almost all GPU memory, at start of stage2. It does check the available memory at the very beginning of the test (when it computes the plan), but allocates it at the start of stage2. So if something reduces the GPU memory at that point, the allocation in stage 2 will fail with a bad_alloc

from gpuowl.

To identify a memory leak, you should observe a gradual reduction in available memory in time. E.g. every time a new P-1 test starts, the GPU available memory is reduced by let's say 100MB, etc.

from gpuowl.

On the failed P-1, what was the "buffers" message at the start of test?
GPU RAM fits 388 stage2 buffers @ 40.0 MB each, using 360

from gpuowl.

• The first exponent:
  GPU RAM fits 184 stage2 buffers @ 40.0 MB each, using 180
• subsequent exponents until now:
  GPU RAM fits 205 stage2 buffers @ 36.0 MB each, using 192

from gpuowl.

{"exponent":"86100473", "worktype":"PM1", "status":"NF", "program":{"name":"gpuowl", "version":"6.4-f6d3153"}, "timestamp":"2019-04-11 15:04:44 UTC", "user":"selroc", "computer":"RX580-9", "fft-length":4718592, "B1":20000, "B2":600000}

verified with:
./openowl -pm1 86100473 -B1 20000 -user selroc -cpu RX580-9 -device 0

No factor found.

from gpuowl.

@valeriob01: That's fine. Here's why:
grep 86100473 test-pm1/pm1.txt
86100473,15290240534639630110561,74,223,323467
So the factor is 15290240534639630110561.
In gp (PARI-GP):
factor(15290240534639630110560)
%1 =
[ 2 5]
[ 3 1]
[ 5 1]
[ 23 1]
[ 223 2]
[ 323467 1]
[86100473 1]

The key is in the exponent of 223; in order to cover 223^2 in B1, we must have B1>=223^2, which is larger than 20000. Try it with B1=50000 or larger.

Thanks for reporting it, luckily it was a bug in the test not in the program :)

from gpuowl.

OK.

./openowl -pm1 86100473 -B1 50000 -user selroc -cpu RX580-9 -device 0

...
2019-04-12 05:57:28 RX580-9 86100473 P-1 GPU RAM fits 205 stage2 buffers @ 36.0 MB each, using 192
...

2019-04-12 06:09:36 RX580-9 {"exponent":"86100473", "worktype":"PM1", "status":"F", "program":{"name":"gpuowl", "version":"6.4-f6d3153"}, "timestamp":"2019-04-12 04:09:36 UTC", "user":"selroc", "computer":"RX580-9", "fft-length":4718592, "B1":50000, "B2":1500000, "factors":["15290240534639630110561"]}

from gpuowl.

Sorry I didn't reply sooner, 4G signal is hard to come by. I meant system memory not GPU memory, this PC just happens to have 16GB of DDR4. Each time a test starts the memory usage of openowl increases when you look at it with top, until it fails to allocate memory for the next test and quits. GPU RAM is unaffected.

from gpuowl.

I think some part of P-1 computation is done on the cpu when a test start or finish because the cpu becomes hot.

from gpuowl.

There might be a leak related to GMP which is used in the GCD computation on CPU. Investigating.

from gpuowl.

Thanks! fixed

from gpuowl.

Wow that was quick. Nice one.

from gpuowl.

from gpuowl.