Hello Stephen,
I've started tests with few angles, with larger and smaller memory
setups, looked at Ignite unit tests and began looking for page eviction
triggers. Once I found interesting code areas I realized that my test
setup was not having proper page eviction set up. This lead me to
further debugging which lead me piece of old code which was overriding
persistence configuration, *effectively* reverting page eviction to
default (no eviction).
After clean up I successfully tested eviction with small caches (512MB
max + direct buffer at 756MB). These worked stable under constant load,
which proved the point that it is possible to have stable operation on
fairly limited resources.
Guilt goes to spring configuration pre/post-processing which went under
radar till today. It was really specific to test application itself, not
Ignite. I thank you for feedback, pointers and assistance!
Kind regards,
Łukasz
On 30.01.2023 12:51, Stephen Darlington wrote:
I’m not saying you should test with 50Gb, but I am saying that 512Mb is
too small. Maybe try a few gigs.
When you define eviction at a cache level, that is for the on-heap
cache. Most people don’t need an on-heap cache.
You you define eviction at a data region level, that works on the
off-heap data. As you observe, it works at a /page/ rather than a record
level. (This is why a tiny region can be problematic. There are far few
pages that can be chosen to evict.)
On 30 Jan 2023, at 11:37, Łukasz Dywicki <l...@code-house.org> wrote:
Dear Jeremy and Stephen,
Thank you for answers. I this issues I face is not bound exclusively
to assigned memory but behavior of eviction policies. After all,
without proper eviction or expiry policy, each amount of memory will
eventually be exceeded. Small size of memory makes it simply easier to
test. To be honest I do not see a point in testing eviction with 50GB
of memory, if it cant be configured for 512MB. Expiry policy proved to
work, but it will not preventing node from crash, as there always
might be high write ratio causing memory to fill.
Both LRU and FIFO eviction policies in default configuration do not
take into account overall memory allocation, they work only with on
heap caches. Unless memory size is explicitly specified they will rely
on over-provisioned memory, or assume that memory assignment is right.
This is possible only if cache entry have a predictable.
Another point is that above eviction policies are dedicated to on-heap
caches and are told to have no affect to off-heap region, why is so?
By looking at sources I see CacheOffheapEvictionManager and
GridCacheEvictionManager, but I can't grasp what enables eviction of
entries from off heap memory managed by Ignite.
Given that my use case is continuous operation with variety of entries
having various sizes I can't assume on-heap cache with LRU nor FIFO,
especially that they do not impact off heap memory. I also can't risk
filling in memory managed by Ignite due to lack of clear eviction
policy during peak load.
As far I can understand from DataRegionConfiguration, its eviction
mechanism works for whole pages, but I haven't seen this happening.
Could it be due to fact that single cache entry in my test exceeds
maximum page size (16MB)?
Best regards,
Łukasz
On 30.01.2023 10:32, Stephen Darlington wrote:
Ignite is designed to work as a cluster with a lot of memory. A
single node with 512Mb of memory just isn’t what most people are
testing with. I’ve seen similar issues when people use tiny nodes. I
don’t see anything immediately wrong with your configuration, but if
it works with more memory, that’s your problem.
On 28 Jan 2023, at 00:35, Łukasz Dywicki <l...@code-house.org> wrote:
Hello again,
I've spent another day on this issue looking at configuration and
with cluster larger than configured backup count ignite did the work.
After that I reverted to testing of single node with minimal
configuration and come to the point that the only way to keep Ignite
survive load was setting ExpiryPolicy to very low value (10s).
To me it seems that Ignite behavior is to preserve all entries in
memory at any cost, even if there is a risk of running into OOM. Is
that true?
I would like to change this behavior and make sure that entries do
not expire because of time as long as there is memory available to
store them. They should be evicted by appropriate EvictionPolicy
only if memory fills up.
To me it looks like DataStoreSettings do not make any impact in this
regard. At least setting page eviction to LRU do not change it.
Please let me know if I am doing something wrong as I can not prove
Ignite to be working stable. Even with such basic objectives I
outlined in earlier mail.
Kind regards,
Łukasz
On 27.01.2023 00:58, Łukasz Dywicki wrote:
Dear all,
I come across use of Apache Ignite to cache results of expensive
computation operation.
Objectives are basic:
- Keep most of "hot" data in memory
- Offload cold part to cache store
- Keep memory utilization under control (evict entries as needed)
While it sounds basic, it doesn't seem to fit Ignite defaults.
What I am testing now is behavior with large objects which can grow
up to 10 mb (serialized) or 25 mb (json representation). Usually
objects will stay far below that threshold, but we can't make
assumption on that.
I began testing various configurations of Ignite in order to
facilitate offloading of memory contents to database. So far I am
stuck for two days at Ignite/application itself running out of
memory after processing several of such large objects. While I know
that storing 10 mb blob in database is not the best idea, I have to
test that behavior too.
By observing database contents I see that number of entries there
grows, but cache do not seem to be evicted. When I try to switch
eviction, it does require onheap to be switched on, and it still
fails with LRU eviction policy.
So far I ended up with a named cache and default region configured
as below:
```
IgniteConfiguration igniteConfiguration = new IgniteConfiguration();
igniteConfiguration.setDataStorageConfiguration(new
DataStorageConfiguration()
.setDefaultDataRegionConfiguration(new DataRegionConfiguration()
.setPersistenceEnabled(false)
.setInitialSize(256L * 1024 * 1024)
.setMaxSize(512L * 1024 * 1024)
.setPageEvictionMode(DataPageEvictionMode.RANDOM_LRU)
.setSwapPath(null)
.setEvictionThreshold(0.75)
)
.setPageSize(DataStorageConfiguration.MAX_PAGE_SIZE)
);
CacheConfiguration<SessionKey, ExpensiveObject> expensiveCache =
new CacheConfiguration<>()
.setName(CACHE_NAME)
.setBackups(2)
.setAtomicityMode(CacheAtomicityMode.ATOMIC)
.setCacheStoreFactory(cacheJdbcBlobStoreFactory)
.setWriteThrough(true)
.setOnheapCacheEnabled(true)
.setEvictionPolicyFactory(new LruEvictionPolicyFactory<>(1024))
.setReadThrough(true);
igniteConfiguration.setCacheConfiguration(
expensiveCache
);
```
What I observe is following - the cache keeps writing data into
database, but it does not remove old entries fast enough to prevent
crash.
JVM parameters I use are fairly basic:
-Xms1g -Xmx1g -XX:+AlwaysPreTouch -XX:+UseG1GC
-XX:+ScavengeBeforeFullGC -XX:+DisableExplicitGC
The store mechanism is jdbc blob store. Exceptions I get happen to
occur in Ignite itself, processing (application code writing cache)
or communication thread used to feed cache. I collected one case here:
https://gist.github.com/splatch/b5ec9134cd9df19bc62f007dd17a19a1
The error message in linked gist advice to enable persistence
(which I did via cache store!), increase memory limit (which I
don't want to do), or enable eviction/expiry policy (which somehow
miss behave).
To me it looks like self defense mechanisms Ignite has are being
tricked leading whole application to crash.
Can you please advise me which settings to tune and how in order to
get Ignite more stable under such load?
Kind regards,
Łukasz
