Segfault in multi-threaded compute node

[thustorage]

Segfault in multi-threaded compute node

[Transpeptidase]

Hi, i found a possible data race bug in the index cache, and pushed a simple solution: 300dc9f. Hope it works

[Transpeptidase]

Hi, is your OS huge page 2MB? If it is 1GB, you should modify the hugepage.sh

[skmonga]

Thanks, that was quick. Yes, the huge page size is 2 MB.

[Transpeptidase]

You can try to reduce the number of huge pages

[skmonga]

I have tried with fewer huge pages as well (4096) but the problem persists.

[Transpeptidase]

It is weird. This issue (#5) reported the same problems, but it results from the huge pages.

[skmonga]

The machine I am using has 128 kB NIC memory, so I have updated Common.h to reflect the same. Is there any possibility it can cause this weird behavior?

[Transpeptidase]

This should not cause a problem...

[Transpeptidase]

Can you comment out this line

Sherman/src/Tree.cpp

Line 674 in 517c96a

index_cache->add_to_cache(page);

to disable index cache ?

[skmonga]

After disabling the index cache, it seems to work fine (no failures on 3 runs of the benchmark). As expected, throughput drops considerably in doing so. Will this be fixed (index cache enabled) for a realistic comparison?

[Transpeptidase]

I don not know what causes this issue.... You can try to comiple with -O0 to figure out if it is a concurrency bugs.

[skmonga]

I am currently testing with -O0 to disable optimizations but the behavior is the same with -O3.

[Transpeptidase]

Hi, can you provide a way to log into your server (my mail is [email protected]) ? In this way, i can debug easily.

[skmonga]

Applying the fix, it works for multi-threaded compute node. Thanks!!