Hi,
> 30 buffers.
JVM kept crashing if I increased memory too much. This was only solution
that gave me lot of memory and did not crash JVM. Not sure why
> 0.2 snapshot
You have outdated documentation then. It shows 0.1 (and fragment names
are wrong BTW).
http://directmemory.apache.org/simple-usage.html
> Serialization
How can I use better serialization?
> JVM settings
I left all at default, except increasing max heap (-Xmx4G for DM,
-Xmx13G for TreeMap).
Are there better settings I should use?
Jan
On 08/11/12 06:44, Raffaele P. Guidi wrote:
Looking at the code I see Jan implemented against v0.1 which is a lot
outdated. I would strongly suggest Jan to use 0.2-SNAPSHOT (yeah, I know
snapshots suck...) and us to work on a new release (it's about time)
focusing a bit more on performance instead of features. What do you guys
think?
Ciao,
R
On Thu, Nov 8, 2012 at 6:16 AM, Raffaele P. Guidi <
[email protected]> wrote:
Oh, sorry I didn't notice the github link (it's 5 'o clock in the morning
here :-D ). Also why thirty 1mb buffers?
Il giorno 08/nov/2012 06:11, "Christoph Engelbert" <[email protected]>
ha scritto:
Unsafe allocator alone won't speed it up that much because I guess he
relied on standard Serialization for his test.
"Raffaele P. Guidi" <[email protected]> schrieb:
I don't believe integers or even 1kb buffers are the right case for
directmemory. Small keys, large payloads (+4kb) are. In any case: how
were
the heap and xxmaxdirectmemory settings? And did you try the Unsafe
storage
as well? It would be helpful if you could share the benchmark code.
Thanks,
R
Il giorno 08/nov/2012 02:49, "Jan Kotek" <[email protected]> ha scritto:
Hi,
MapDB now has DirectMemory storage.
I made little test to compare our two libraries + java.util.TreeMap
All tests are on 64bit Linux, 16GB RAM, JDK6. Default JVM settings.
Source code is here: https://github.com/jankotek/**
mapdb-junk/tree/master/src/**main/java/**directMemoryBenchmark<
https://github.com/jankotek/mapdb-junk/tree/master/src/main/java/directMemoryBenchmark
First is performance test. How long does it take to insert and fetch
1
million records, each 1KB large.
- DM - 6901 ms
- MapDB - 2045 ms
- java.util.TreeMap - 940 ms
Second is overhead test. It inserts pair of integers in cycle, until
JVM
runs out of memory.
It should measure how much memory library each consumes for its own
structures.
- DM - inserts 9 000 000 records and then it becomes very slow.
Profiler
shows high GC activity.
- MapDB - works about 10 minutes and than it runs out of memory with
217
000 000 records.
- java.util.TreeMap - crashes after 99 000 000 records , it consumed
about 12GB heap.
I am not sure I configured DM correctly, JVM kept crashing when I
increased size too much.
I ended with this setting:
CacheService<Object, Object> m = new DirectMemory<Object,
Object>()
.setNumberOfBuffers( 30 )
.setSize((int) 1e8)
.setInitialCapacity( 10000 )
.setConcurrencyLevel( 4 )
.setDisposalTime(1000000)
.newCacheService();
Hope this helps.
Jan
--
Diese Nachricht wurde von meinem Android-Mobiltelefon mit K-9 Mail
gesendet.
