[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13920599#comment-13920599 ] ramkrishna.s.vasudevan edited comment on HBASE-10191 at 3/5/14 7:53 AM: bq.Would be sweet if the value at least was not on heap Yes, this could be a nice one. So I think before doing this the usage of Cell should be in place. {Got added by mistake.} was (Author: ram_krish): bq.Would be sweet if the value at least was not on heap Yes, this could be a nice. > Move large arena storage off heap > - > > Key: HBASE-10191 > URL: https://issues.apache.org/jira/browse/HBASE-10191 > Project: HBase > Issue Type: Umbrella >Reporter: Andrew Purtell > > Even with the improved G1 GC in Java 7, Java processes that want to address > large regions of memory while also providing low high-percentile latencies > continue to be challenged. Fundamentally, a Java server process that has high > data throughput and also tight latency SLAs will be stymied by the fact that > the JVM does not provide a fully concurrent collector. There is simply not > enough throughput to copy data during GC under safepoint (all application > threads suspended) within available time bounds. This is increasingly an > issue for HBase users operating under dual pressures: 1. tight response SLAs, > 2. the increasing amount of RAM available in "commodity" server > configurations, because GC load is roughly proportional to heap size. > We can address this using parallel strategies. We should talk with the Java > platform developer community about the possibility of a fully concurrent > collector appearing in OpenJDK somehow. Set aside the question of if this is > too little too late, if one becomes available the benefit will be immediate > though subject to qualification for production, and transparent in terms of > code changes. However in the meantime we need an answer for Java versions > already in production. This requires we move the large arena allocations off > heap, those being the blockcache and memstore. On other JIRAs recently there > has been related discussion about combining the blockcache and memstore > (HBASE-9399) and on flushing memstore into blockcache (HBASE-5311), which is > related work. We should build off heap allocation for memstore and > blockcache, perhaps a unified pool for both, and plumb through zero copy > direct access to these allocations (via direct buffers) through the read and > write I/O paths. This may require the construction of classes that provide > object views over data contained within direct buffers. This is something > else we could talk with the Java platform developer community about - it > could be possible to provide language level object views over off heap > memory, on heap objects could hold references to objects backed by off heap > memory but not vice versa, maybe facilitated by new intrinsics in Unsafe. > Again we need an answer for today also. We should investigate what existing > libraries may be available in this regard. Key will be avoiding > marshalling/unmarshalling costs. At most we should be copying primitives out > of the direct buffers to register or stack locations until finally copying > data to construct protobuf Messages. A related issue there is HBASE-9794, > which proposes scatter-gather access to KeyValues when constructing RPC > messages. We should see how far we can get with that and also zero copy > construction of protobuf Messages backed by direct buffer allocations. Some > amount of native code may be required. -- This message was sent by Atlassian JIRA (v6.2#6252)
[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13917703#comment-13917703 ] Andrew Purtell edited comment on HBASE-10191 at 3/3/14 2:56 AM: bq. The problem is that if you have hundreds of 1MB in-memory HFiles, then it becomes too expensive to merge them all (via KVHeap) when scanning. A possible solution is to subdivide the memstore into "stripes" (probably smaller than the stripe compaction stripes) and periodically compact the in-memory stripes Anoop, Ram, and I were throwing around ideas of making in-memory HFiles out of memstore snapshots, and then doing in-memory compaction over them. If we have off-heap backing for memstore we could potentially carry larger snapshots (in memory HFiles resulting from a few merged memstore snapshots) leading to less frequent flushes and significantly less write amplification overall. was (Author: apurtell): bq. The problem is that if you have hundreds of 1MB in-memory HFiles, then it becomes too expensive to merge them all (via KVHeap) when scanning. A possible solution is to subdivide the memstore into "stripes" (probably smaller than the stripe compaction stripes) and periodically compact the in-memory stripes Anoop, Ram, and I were throwing around ideas of making in-memory HFiles out of memstore snapshots, and then doing in-memory compaction over them. If we have off-heap backing for memstore we could potentially carry larger datasets leading to less frequent flushes and significantly less write amplification overall. > Move large arena storage off heap > - > > Key: HBASE-10191 > URL: https://issues.apache.org/jira/browse/HBASE-10191 > Project: HBase > Issue Type: Umbrella >Reporter: Andrew Purtell > > Even with the improved G1 GC in Java 7, Java processes that want to address > large regions of memory while also providing low high-percentile latencies > continue to be challenged. Fundamentally, a Java server process that has high > data throughput and also tight latency SLAs will be stymied by the fact that > the JVM does not provide a fully concurrent collector. There is simply not > enough throughput to copy data during GC under safepoint (all application > threads suspended) within available time bounds. This is increasingly an > issue for HBase users operating under dual pressures: 1. tight response SLAs, > 2. the increasing amount of RAM available in "commodity" server > configurations, because GC load is roughly proportional to heap size. > We can address this using parallel strategies. We should talk with the Java > platform developer community about the possibility of a fully concurrent > collector appearing in OpenJDK somehow. Set aside the question of if this is > too little too late, if one becomes available the benefit will be immediate > though subject to qualification for production, and transparent in terms of > code changes. However in the meantime we need an answer for Java versions > already in production. This requires we move the large arena allocations off > heap, those being the blockcache and memstore. On other JIRAs recently there > has been related discussion about combining the blockcache and memstore > (HBASE-9399) and on flushing memstore into blockcache (HBASE-5311), which is > related work. We should build off heap allocation for memstore and > blockcache, perhaps a unified pool for both, and plumb through zero copy > direct access to these allocations (via direct buffers) through the read and > write I/O paths. This may require the construction of classes that provide > object views over data contained within direct buffers. This is something > else we could talk with the Java platform developer community about - it > could be possible to provide language level object views over off heap > memory, on heap objects could hold references to objects backed by off heap > memory but not vice versa, maybe facilitated by new intrinsics in Unsafe. > Again we need an answer for today also. We should investigate what existing > libraries may be available in this regard. Key will be avoiding > marshalling/unmarshalling costs. At most we should be copying primitives out > of the direct buffers to register or stack locations until finally copying > data to construct protobuf Messages. A related issue there is HBASE-9794, > which proposes scatter-gather access to KeyValues when constructing RPC > messages. We should see how far we can get with that and also zero copy > construction of protobuf Messages backed by direct buffer allocations. Some > amount of native code may be required. -- This message was sent by Atlassian JIRA (v6.2#6252)
[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13906599#comment-13906599 ] Lars Hofhansl edited comment on HBASE-10191 at 2/20/14 4:33 PM: My office neighbor used to work on a proprietary Java database, and he says they used 128GB or even 192GB Java heaps and larger all the time without any significant GC impact. (non moving) Collection times are not a function of the heap size but rather of heap complexity, i.e. the number of objects to track (HBase also produces a lot of garbage, but that is short lived and can be quickly collected by a moving collector for the young gen). With memstoreLAB and the block cache HBase already does a good job on this. Even as is currently, if we fill an entire 128GB of heap with 64k blocks from the blockcache that would only be about 2m objects. Now, if we want < 100ms latency area we need to rethink things; that will generally be very difficult in current Java. While we move all-or-nothing everything out of the Java heap, we should also investigate whether we can make the GC's life easier, yet. Edit: Edited for clarity. was (Author: lhofhansl): This might not be very popular viewpoint these days, but anyway. My office neighbor used to work on a proprietary Java database, and he says they used 128GB or even 192GB Java heaps and larger all the time without any significant GC impact. (non moving) Collection times are not a function of the heap size but rather of heap complexity, i.e. the number of objects to track (HBase also produces a lot of garbage, but that is short lived and can be quickly collected by a moving collector for the young gen). With memstoreLAB and the block cache HBase already does a good job on this. Even as is currently, if we fill an entire 128GB of heap with 64k blocks from the blockcache that would only be about 2m objects. Now, if we want to forage into the < 100ms latency area we need to rethink things, but then Java might just not be the right choice. Before we embark on an all-or-nothing adventure and move everything out of the Java heap, we should also investigate whether we can make the GC's life easier, yet. > Move large arena storage off heap > - > > Key: HBASE-10191 > URL: https://issues.apache.org/jira/browse/HBASE-10191 > Project: HBase > Issue Type: Umbrella >Reporter: Andrew Purtell > > Even with the improved G1 GC in Java 7, Java processes that want to address > large regions of memory while also providing low high-percentile latencies > continue to be challenged. Fundamentally, a Java server process that has high > data throughput and also tight latency SLAs will be stymied by the fact that > the JVM does not provide a fully concurrent collector. There is simply not > enough throughput to copy data during GC under safepoint (all application > threads suspended) within available time bounds. This is increasingly an > issue for HBase users operating under dual pressures: 1. tight response SLAs, > 2. the increasing amount of RAM available in "commodity" server > configurations, because GC load is roughly proportional to heap size. > We can address this using parallel strategies. We should talk with the Java > platform developer community about the possibility of a fully concurrent > collector appearing in OpenJDK somehow. Set aside the question of if this is > too little too late, if one becomes available the benefit will be immediate > though subject to qualification for production, and transparent in terms of > code changes. However in the meantime we need an answer for Java versions > already in production. This requires we move the large arena allocations off > heap, those being the blockcache and memstore. On other JIRAs recently there > has been related discussion about combining the blockcache and memstore > (HBASE-9399) and on flushing memstore into blockcache (HBASE-5311), which is > related work. We should build off heap allocation for memstore and > blockcache, perhaps a unified pool for both, and plumb through zero copy > direct access to these allocations (via direct buffers) through the read and > write I/O paths. This may require the construction of classes that provide > object views over data contained within direct buffers. This is something > else we could talk with the Java platform developer community about - it > could be possible to provide language level object views over off heap > memory, on heap objects could hold references to objects backed by off heap > memory but not vice versa, maybe facilitated by new intrinsics in Unsafe. > Again we need an answer for today also. We should investigate what existing > libraries may be available in this regard. Key will be avoiding > marshalling/unmarshalling
[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13906456#comment-13906456 ] Andrew Purtell edited comment on HBASE-10191 at 2/20/14 2:07 AM: - I'm looking at Netty 4's netty-buffer module (http://netty.io/4.0/api/io/netty/buffer/package-summary.html), which has some nice properties, including composite buffers, arena allocation, dynamic buffer resizing, and reference counting, never mind dev and testing by another community. I also like it because you can plug in your own allocators and specialize the abstract ByteBuf base type. More on this later. When I get closer to seeing what exactly needs to be done I will post a design doc. Current thinking follows. Below the term 'buffer' currently means Netty ByteBufs or derived classes backed by off-heap allocated direct buffers. *Write* When coming in from RPC, cells are laid out by codecs into cellbocks in buffers and the cellblocks/buffers are handed to the memstore. Netty's allocation arenas replace the MemstoreLAB. The memstore data structure evolves into an index over cellblocks. Per [~mcorgan]'s comment above, we should think about how the memstore index can be built with fewer object allocations than the number of cells in the memstore, yet be in the ballpark with efficiency of concurrent access. A tall order. CSLM wouldn't be the right choice as it allocates at least one list entry per key, but we could punt and use it initially and make a replacement datastructure as a follow on task. Cellblocks in memstore should be amenable to flushing to disk as a gathering write. This may mean cellblocks have the same internal structure as HFile blocks and we reuse all of the block encoder machinery (and simplify them in the process). *Read* We feed down buffers to HDFS to fill with file block data. We pick which pool to get a buffer from for a read depending on family caching strategy. Pools could be backed by arenas that match up with LRU policy strata, with a common pool/arena for noncaching reads. (Or for noncaching reads, can we optionally use a new API for getting buffers up from HDFS, perhaps backed by the pinned shared RAM cache, since we know we will be referring to the contents only briefly?) It will be important to get reference counting right as we will be servicing scans while attempting to evict. Related, eviction of a block may not immediately return a buffer to a pool, if there is more than one block in a buffer. We maintain new metrics on numbers of buffers allocated, stats on arenas, stats on wastage and internal fragmentation of the buffers, etc, and use these to guide optimizations and refinements. This should require fewer changes than the write side since we are already set up for dealing with cellblocks. Design points to optimize would be minimizing the number and size of data copies, minimizing the number of on-heap object allocations, and on disk encoding suitable as an efficient in-memory representation. was (Author: apurtell): I'm looking at Netty 4's netty-buffer module (http://netty.io/4.0/api/io/netty/buffer/package-summary.html), which has some nice properties, including composite buffers, arena allocation, dynamic buffer resizing, and reference counting, never mind dev and testing by another community. I also like it because you can plug in your own allocators and specialize the abstract ByteBuf base type. More on this later. When I get closer to seeing what exactly needs to be done I will post a design doc. Current thinking follows. Below the term 'buffer' currently means Netty ByteBufs or derived classes backed by off-heap allocated direct buffers. *Write* When coming in from RPC, cells are laid out by codecs into cellbocks in buffers and the cellblocks/buffers are handed to the memstore. Netty's allocation arenas replace the MemstoreLAB. The memstore data structure evolves into an index over cellblocks. Per [~mcorgan]'s comment above, we should think about how the memstore index can be built with fewer object allocations than the number of cells in the memstore, yet be in the ballpark with efficiency of concurrent access. A tall order. CSLM wouldn't be the right choice as it allocates at least one list entry per key, but we could punt and use it initially and make a replacement datastructure as a follow on task. Cellblocks in memstore should be amenable to flushing to disk as a gathering write. This may mean cellblocks have the same internal structure as HFile blocks and we reuse all of the block encoder machinery (and simplify them in the process). *Read* We feed down buffers to HDFS to fill with file block data. We pick which pool to get a buffer from for a read depending on family caching strategy. Pools could be backed by arenas that match up with LRU policy strata, with a common pool/arena for
[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13906456#comment-13906456 ] Andrew Purtell edited comment on HBASE-10191 at 2/20/14 1:59 AM: - I'm looking at Netty 4's netty-buffer module (http://netty.io/4.0/api/io/netty/buffer/package-summary.html), which has some nice properties, including composite buffers, arena allocation, dynamic buffer resizing, and reference counting, never mind dev and testing by another community. I also like it because you can plug in your own allocators and specialize the abstract ByteBuf base type. More on this later. When I get closer to seeing what exactly needs to be done I will post a design doc. Current thinking follows. Below the term 'buffer' currently means Netty ByteBufs or derived classes backed by off-heap allocated direct buffers. *Write* When coming in from RPC, cells are laid out by codecs into cellbocks in buffers and the cellblocks/buffers are handed to the memstore. Netty's allocation arenas replace the MemstoreLAB. The memstore data structure evolves into an index over cellblocks. Per [~mcorgan]'s comment above, we should think about how the memstore index can be built with fewer object allocations than the number of cells in the memstore, yet be in the ballpark with efficiency of concurrent access. A tall order. CSLM wouldn't be the right choice as it allocates at least one list entry per key, but we could punt and use it initially and make a replacement datastructure as a follow on task. Cellblocks in memstore should be amenable to flushing to disk as a gathering write. This may mean cellblocks have the same internal structure as HFile blocks and we reuse all of the block encoder machinery (and simplify them in the process). *Read* We feed down buffers to HDFS to fill with file block data. We pick which pool to get a buffer from for a read depending on family caching strategy. Pools could be backed by arenas that match up with LRU policy strata, with a common pool/arena for noncaching reads. (Or for noncaching reads, can we optionally use a new API for getting buffers up from HDFS, perhaps backed by the pinned shared RAM cache, since we know we will be referring to the contents only briefly?) It will be important to get reference counting right as we will be servicing scans while attempting to evict. Related, eviction of a block may not immediately return a buffer to a pool, if there is more than one block in a buffer. We maintain new metrics on numbers of buffers allocated, stats on arenas, stats on wastage and internal fragmentation of the buffers, etc, and use these to guide optimizations and refinements. was (Author: apurtell): I'm looking at Netty 4's netty-buffer module (http://netty.io/4.0/api/io/netty/buffer/package-summary.html), which has some nice properties, including composite buffers, arena allocation, dynamic buffer resizing, and reference counting, never mind dev and testing by another community. I also like it because you can plug in your own allocators and specialize the abstract ByteBuf base type. More on this later. When I get closer to seeing what exactly needs to be done I will post a design doc. Current thinking follows. Below the term 'buffer' currently means Netty ByteBufs or derived classes backed by off-heap allocated direct buffers. *Write* When coming in from RPC, cells are laid out by codecs into cellbocks in buffers and the cellblocks/buffers are handed to the memstore. Netty's allocation arenas replace the MemstoreLAB. The memstore data structure evolves into an index over cellblocks. Per [~mcorgan]'s comment above, we should think about how the memstore index can be built with fewer object allocations than the number of cells in the memstore, yet be in the ballpark with efficiency of concurrent access. A tall order. CSLM wouldn't be the right choice as it allocates at least one list entry per key, but we could punt and use it initially and make a replacement datastructure as a follow on task. *Read* We feed down buffers to HDFS to fill with file block data. We pick which pool to get a buffer from for a read depending on family caching strategy. Pools could be backed by arenas that match up with LRU policy strata, with a common pool/arena for noncaching reads. (Or for noncaching reads, can we optionally use a new API for getting buffers up from HDFS, perhaps backed by the pinned shared RAM cache, since we know we will be referring to the contents only briefly?) It will be important to get reference counting right as we will be servicing scans while attempting to evict. Related, eviction of a block may not immediately return a buffer to a pool, if there is more than one block in a buffer. We maintain new metrics on numbers of buffers allocated, stats on arenas, stats on wastage and internal fra
[jira] [Comment Edited] (HBASE-10191) Move large arena storage off heap
[ https://issues.apache.org/jira/browse/HBASE-10191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13851148#comment-13851148 ] Andrew Purtell edited comment on HBASE-10191 at 12/18/13 12:18 AM: --- bq. Memstore and BlockCache are commonly cited as the offending components, but I've not seen anyone present conclusive profiling results making this clear It's abundantly clear once using heaps larger than ~8 GB that collection pauses under safepoint blow out latency SLAs at the high percentiles. Why would we need heaps larger than this? To take direct advantage of large server RAM. Memstore and blockcache are then the largest allocators of heap memory. If we move them off heap, they can "soak up" most of the available RAM, leaving remaining heap demand relatively small - this is the idea. Edit: Phrasing was (Author: apurtell): bq. Memstore and BlockCache are commonly cited as the offending components, but I've not seen anyone present conclusive profiling results making this clear It's abundantly clear once using heaps larger than ~8 GB that collection pauses under safepoint blow out latency SLAs at the high percentiles. I've observed this directly under mixed read+write load. (Read-only loads work ok with G1 even with very large heaps, e.g. 192 GB.) Why would we need heaps larger than this? To take direct advantage of large server RAM. Memstore and blockcache are then the largest allocators of heap memory. If we move them off heap, they can "soak up" most of the available RAM, leaving remaining heap demand relatively small - this is the idea. > Move large arena storage off heap > - > > Key: HBASE-10191 > URL: https://issues.apache.org/jira/browse/HBASE-10191 > Project: HBase > Issue Type: Umbrella >Reporter: Andrew Purtell > > Umbrella issue for moving large arena storage off heap. > Even with the improved G1 GC in Java 7, Java processes that want to address > large regions of memory while also providing low high-percentile latencies > continue to be challenged. Fundamentally, a Java server process that has high > data throughput and also tight latency SLAs will be stymied by the fact that > the JVM does not provide a fully concurrent collector. There is simply not > enough throughput to copy data during GC under safepoint (all application > threads suspended) within available time bounds. This is increasingly an > issue for HBase users operating under dual pressures: 1. tight response SLAs, > 2. the increasing amount of RAM available in "commodity" server > configurations, because GC load is roughly proportional to heap size. > We can address this using parallel strategies. We should talk with the Java > platform developer community about the possibility of a fully concurrent > collector appearing in OpenJDK somehow. Set aside the question of if this is > too little too late, if one becomes available the benefit will be immediate > though subject to qualification for production, and transparent in terms of > code changes. However in the meantime we need an answer for Java versions > already in production. This requires we move the large arena allocations off > heap, those being the blockcache and memstore. On other JIRAs recently there > has been related discussion about combining the blockcache and memstore > (HBASE-9399) and on flushing memstore into blockcache (HBASE-5311), which is > related work. We should build off heap allocation for memstore and > blockcache, perhaps a unified pool for both, and plumb through zero copy > direct access to these allocations (via direct buffers) through the read and > write I/O paths. This may require the construction of classes that provide > object views over data contained within direct buffers. This is something > else we could talk with the Java platform developer community about - it > could be possible to provide language level object views over off heap > memory, on heap objects could hold references to objects backed by off heap > memory but not vice versa, maybe facilitated by new intrinsics in Unsafe. > Again we need an answer for today also. We should investigate what existing > libraries may be available in this regard. Key will be avoiding > marshalling/unmarshalling costs. At most we should be copying primitives out > of the direct buffers to register or stack locations until finally copying > data to construct protobuf Messages. A related issue there is HBASE-9794, > which proposes scatter-gather access to KeyValues when constructing RPC > messages. We should see how far we can get with that and also zero copy > construction of protobuf Messages backed by direct buffer allocations. Some > amount of native code may be required. -- This message was sent by Atlassian JIRA (v6.1.4#6159)