Problems with increasing/decreasing cache size when live

Question

Hello,
I have configured multiple environments which I'm compacting sequentially and to achieve this I allocate a bigger cache to the env currently being compacted as follows:
Initialization:
DB_ENV->set_cachesize(gbytes, bytes, 1); // Initial cache size.
DB_ENV->set_cache_max(gbytes, bytes); // Maximum size.
While live, application decreases cache of current env when finished and then increases cache of next env using:
DB_ENV->set_cachesize(gbytes, obytes, 0); // Decrease cache size of current env to initial size
DB_ENV->set_cachesize(gbytes, obytes, 0); // Increase cache size of next env to max size.
When I print statistics about the memory pool using DB_ENV->memp_stat I can see that everyting is going normally:
memp_stat: env1 ncache= 8 cache_size=20973592 // env1 is current env
memp_stat: env2 ncache= 1 cache_size=20973592
and then after changing current env:
memp_stat: env1 ncache= 1 cache_size=20973592
memp_stat: env2 ncache= 8 cache_size=20973592 // env2 is now current env
But the problem is that over time memory is leaked (as if the extra memory of each env was not freed) and I'm totally sure that the problem comes from this code.
I'm running Berkeley DB 4.7.25 on FreeBSD.
Maybe some leak was fixed in newer versions and you could suggest to me a patch? or I don't use the API correctly?
Thanks!
Edited by: 894962 on Jan 23, 2012 6:40 AM

Answer 1

Hi,
Thanks for providing the information.
Unfortunately, I don't remember exact test case I was doing, so I did a new one with 32 env.
I set the following for each env:
- Initial cache=512MB/32
- Max=1GB
Before open, I do:
DBEnvironment->set_cachesize((u_int32_t)0, (u_int32_t)512*1024*1024/32, 1);
DBEnvironment->set_cache_max(1*1024*1024*1024, 0);
DBEnvironment->get_cache_max(&gbytes, &obytes); // gives gbytes=1 and obytes=0
After open, I have the following:
DBEnvironment->get_cache_max(&gbytes, &obytes); // gives gbytes=0 and obytes=9502720
memp_stat: cache_size=18644992 cache_ncache=1
So here, the values returned by memp_stat are normal but get_cache_max is strange. Then after increasing the cache to the strange value returned by get_cache_max (gbytes=0, obytes=9502720), I have the following:
DBEnvironment->get_cache_max(&gbytes, &obytes); // gives gbytes=0 and obytes=9502720
memp_stat: outlinks cache_size=27328512 cache_ncache=54
with cache_size being: ((ui64)sp->st_gbytes * GIGA + sp->st_bytes);.
So cache is actually increased...
I try to reproduce this case by opening 1 env as follows.
//Before open
DbEnv->set_cachesize(); 512MB, 1 cache
DbEnv->set_cache_max; 1GB
//After open
DbEnv->get_cachesize; 512MB, 1cache
DbEnv->get_caceh_max; 1GB
memp_stat: cache:512MB, ncache:1, cache_max:1GB
//Decrease the cache size
DbEnv->set_cachesize(); 9MB(9502720B), 1 cache
DbEnv->get_cachesize; 512MB, 1cache
DbEnv->get_caceh_max; 1GB
memp_stat: cache:512MB, ncache:1, cache_max:1GB
All the result is expected. Since when resizing the cache after DbEnv is open, it is rounded to the nearest multiple of the region size. Region size means the size of each region specified initially. Please refer to BDB doc: [http://docs.oracle.com/cd/E17076_02/html/api_reference/C/envset_cachesize.html|http://docs.oracle.com/cd/E17076_02/html/api_reference/C/envset_cachesize.html]. Here region size is 512MB/1cache = 512MB. And I don't think you can resize the cache smaller than 1 region.
Since you are opening 32 env at the same time with 512MB cache and 1GB maximum for each, when the env is open, whether it can allocate as much as that specified for the cache, is dependent on the system. I am guess the number 9502720 got from get_cache_max after opening the env is probably based on the system and app request, the cache size you can get when opening the env.
And for the case listed in the beginning of the post
While live, application decreases cache of current env when finished and then increases cache of next env using:
DB_ENV->set_cachesize(gbytes, obytes, 0); // Decrease cache size of current env to initial size
DB_ENV->set_cachesize(gbytes, obytes, 0); // Increase cache size of next env to max size.
When I print statistics about the memory pool using DB_ENV->memp_stat I can see that everyting is going normally:
memp_stat: env1 ncache= 8 cache_size=20973592 // env1 is current env
memp_stat: env2 ncache= 1 cache_size=20973592
and then after changing current env:
memp_stat: env1 ncache= 1 cache_size=20973592
memp_stat: env2 ncache= 8 cache_size=20973592 // env2 is now current envWhen env1 is finishing soon, what numbers do you set in set_cachesize to decrease the cache, including the number of caches and cache size?
When decreasing the cache, I do:
env->GetDbEnv()->set_cachesize((u_int32_t)0, (u_int32_t)20973592, 0);
I mean, in all cases I simply set cachesize to its original value (obtained after open through get_cachesize) when decreasing and set cachesize to its max value when increasing (obtained though get_cache_max; plus I do something like cacheMaxSize * 0.75 if < 500MB).I can reproduce this case. And I think the result is expected. When using DBEnv->set_cachesize() to resize the cache after env is opened, the ncache para is ignored. Please refer to BDB doc here: [http://docs.oracle.com/cd/E17076_02/html/api_reference/C/envset_cachesize.html|http://docs.oracle.com/cd/E17076_02/html/api_reference/C/envset_cachesize.html] . Hence I don't think you can decrease the cache size by setting the number of cache to 0.
Hope it helps.