[
https://issues.apache.org/jira/browse/HBASE-4027?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13073097#comment-13073097
]
[email protected] commented on HBASE-4027:
------------------------------------------------------
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > conf/hbase-env.sh, lines 44-45
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28125#file28125line44>
bq. >
bq. > Is MaxDirectMemorySize determinable on the running jvm? Could we
make the offheapcachesize config as a percentage of the direct memory size like
we have for memstore/blockcache today? (default of 0.95 or something would
make it so it never really has to be set for most cases... and i'm not sure
what exactly "a bit above the off heap cache size" is)
I haven't figured out a way. Just asked StackOverflow - hopefully they'll have
an answer.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCacheTestUtils.java,
line 2
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28127#file28127line2>
bq. >
bq. > 2011
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/CacheStats.java, line 1
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28128#file28128line1>
bq. >
bq. > license
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCacheTestUtils.java,
line 41
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28127#file28127line41>
bq. >
bq. > whitespace
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/CacheStats.java, line 5
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28128#file28128line5>
bq. >
bq. > class comment
Added
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java,
line 39
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28129#file28129line39>
bq. >
bq. > Would it make sense to have DoubleBlockCache be more generic? Does
it need to be fixed with these two types or could it take two BlockCache's and
they are executed in the order they are given in (just need to be clear in doc).
bq. >
bq. > If this was generic, it could be reused for various multi-level
caches (like an underlying cache with compressed blocks and one above it with
uncompressed blocks)
It can be made more generic easily. I just haven't done it. Will do though -
just create a constructor that takes two BlockCaches?
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java,
line 66
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28129#file28129line66>
bq. >
bq. > longer than 80 chars
fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java,
lines 77-79
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28129#file28129line77>
bq. >
bq. > This seems like a behavior that we may not always want.
bq. >
bq. > If we made this class generic, could we have some kind of policy we
initiate it with? (like default cache in level one, if accessed in level one,
cache in level two, etc?)
bq. >
bq. > we're going to always be double-storing anything so that the offHeap
true capacity is (totalOffHeap - totalOnHeap). in some cases, we might want to
cache on heap first and then if evicted we cache off heap, or maybe we want it
to work more like the existing LRU (first read goes into off heap, second read
upgrades it to the on heap cache and removes from the off heap)
I was thinking of sending all initial caches to the off heap cache, giving it
an inbuilt scan resistance, but LRUBlockCache already does partitioning to deal
with scans.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java,
line 115
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28129#file28129line115>
bq. >
bq. > this is going to make for some weird stats? seems like we may need
to actually expose the stats of each underlying cache rather than both? (or
both and separate). it's going to be difficult to understand what's happening
when the hit and eviction stats cover both.
The idea was the combined evicted stats of both, but yeah, CacheStats are
implemented for both LruBlockCache and SlabCache, so exposing both might be a
good idea.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SingleSizeCache.java, line
74
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28132#file28132line74>
bq. >
bq. > huh?
Just a typo. Fixed. Don't know why that line is there.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SingleSizeCache.java, line
84
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28132#file28132line84>
bq. >
bq. > line > 80 chars
fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java,
lines 122-125
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line122>
bq. >
bq. > seems like an odd behavior. this method is kinda non-deterministic
from the caller POV, we have no idea whether this passed or failed
If the size of the item being cached is too big to fit into any slab, we can't
cache it. Again, I'm moving away from MemCached's model, so this should work
better.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java, line
151
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line151>
bq. >
bq. > is this really a cache miss? this cache will never take this block.
Good point. Taking that part out.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java, line
155
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line155>
bq. >
bq. > but here we need to check for null and increment cache miss rather
than hit if it is null
Yup, doing so as well.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java,
lines 170-173
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line170>
bq. >
bq. > this non-javadoc comment is not necessary, just the @Override has
been normal format in hbase
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java, line
269
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line269>
bq. >
bq. > should this really be a bunch of INFO logs? Could we have these
exported as metrics instead?
Will take care of this next pass.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabItemEvictionWatcher.java,
line 1
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28135#file28135line1>
bq. >
bq. > license
added.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/regionserver/StoreFile.java, lines
374-375
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28137#file28137line374>
bq. >
bq. > formatting looks a little strange here (missing spaces and such)
fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSingleSlabCache.java,
line 28
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28138#file28138line28>
bq. >
bq. > This test looks more specific than testing "the concurrent
LruBlockCache"
Woah, typo, fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSingleSlabCache.java,
line 34
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28138#file28138line34>
bq. >
bq. > you are missing a space here and throughout this file and some
others... our format is to always have a space before a {
fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSlab.java, line 1
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28139#file28139line1>
bq. >
bq. > license
added.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. >
src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSingleSlabCache.java,
line 2
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28138#file28138line2>
bq. >
bq. > 2011
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/regionserver/StoreFile.java, line
376
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28137#file28137line376>
bq. >
bq. > line > 80 chars
Fixed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java,
lines 63-65
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line63>
bq. >
bq. > Why these ratios? At the least, this should all be configurable
(even if just in code and undocumented).
bq. >
bq. > Do we need to always pre-allocate everything and determine the
block/slab sizes and all that? The design seems inflexible because it's all
determine during construction rather than being adaptive.
bq. >
bq. > I'm okay with the first iteration not being awesome and auto-tuning
but this layered cache design seems to make it hard to change anything once
it's instantiated.
Its sort of expensive to change cache sizes with DirectByteBuffer after
instantiation. I have a plan to restructure this away from MemCached's model
though.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java,
line 37
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28129#file28129line37>
bq. >
bq. > whitespace here and throughout this file
Removed.
bq. On 2011-07-29 17:41:50, Jonathan Gray wrote:
bq. > src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java, line
48
bq. > <https://reviews.apache.org/r/1214/diff/1/?file=28134#file28134line48>
bq. >
bq. > I'm not totally clear on why the SlabCache contains a bunch of
SingleSizeCaches. Why do you need to layer BlockCaches on top of BlockCaches?
You'll have one slab per size rather than one cache per size? Can you not pass
the right evictor callback in so it goes back to the right slab?
I'm using MemCached's allocation model - although I have one written that does
it in a wholly different manner. But MemCached gets around fragmentation by
allocating a certain range of sizes to a Single Slab.
If something is smaller than the max size of a particular slab, the extra space
is just wasted. Each SingleSizeCache can cache things up to a certain size.
Things above that size cannot be cached, and things below that size will waste
space.
- Li
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/1214/#review1211
-----------------------------------------------------------
On 2011-07-30 00:39:48, Li Pi wrote:
bq.
bq. -----------------------------------------------------------
bq. This is an automatically generated e-mail. To reply, visit:
bq. https://reviews.apache.org/r/1214/
bq. -----------------------------------------------------------
bq.
bq. (Updated 2011-07-30 00:39:48)
bq.
bq.
bq. Review request for hbase, Todd Lipcon, Ted Yu, Michael Stack, Jonathan
Gray, and Li Pi.
bq.
bq.
bq. Summary
bq. -------
bq.
bq. Review request - I apparently can't edit tlipcon's earlier posting of my
diff, so creating a new one.
bq.
bq.
bq. This addresses bug HBase-4027.
bq. https://issues.apache.org/jira/browse/HBase-4027
bq.
bq.
bq. Diffs
bq. -----
bq.
bq. conf/hbase-env.sh 2d55d27
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCache.java 509121d
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCacheTestUtils.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/CacheStats.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/DoubleBlockCache.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/LruBlockCache.java
ecab7ca
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/SimpleBlockCache.java
150f54f
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SingleSizeCache.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/slab/Slab.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabCache.java
PRE-CREATION
bq.
src/main/java/org/apache/hadoop/hbase/io/hfile/slab/SlabItemEvictionWatcher.java
PRE-CREATION
bq. src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java
1d5e3fa
bq. src/main/java/org/apache/hadoop/hbase/regionserver/StoreFile.java
b600020
bq.
src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSingleSlabCache.java
PRE-CREATION
bq. src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSlab.java
PRE-CREATION
bq. src/test/java/org/apache/hadoop/hbase/io/hfile/slab/TestSlabCache.java
PRE-CREATION
bq.
bq. Diff: https://reviews.apache.org/r/1214/diff
bq.
bq.
bq. Testing
bq. -------
bq.
bq. Ran benchmarks against it in HBase standalone mode. Wrote test cases for
all classes, multithreaded test cases exist for the cache.
bq.
bq.
bq. Thanks,
bq.
bq. Li
bq.
bq.
> Enable direct byte buffers LruBlockCache
> ----------------------------------------
>
> Key: HBASE-4027
> URL: https://issues.apache.org/jira/browse/HBASE-4027
> Project: HBase
> Issue Type: Improvement
> Reporter: Jason Rutherglen
> Assignee: Li Pi
> Priority: Minor
> Attachments: 4027-v5.diff, 4027v7.diff, HBase-4027 (1).pdf,
> HBase-4027.pdf, hbase-4027v6.diff, slabcachepatch.diff,
> slabcachepatchv2.diff, slabcachepatchv3.1.diff, slabcachepatchv3.2.diff,
> slabcachepatchv3.diff, slabcachepatchv4.5.diff, slabcachepatchv4.diff
>
>
> Java offers the creation of direct byte buffers which are allocated outside
> of the heap.
> They need to be manually free'd, which can be accomplished using an
> documented {{clean}} method.
> The feature will be optional. After implementing, we can benchmark for
> differences in speed and garbage collection observances.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
