[jira] [Comment Edited] (HIVE-20380) explore storing multiple CBs in a single cache buffer in LLAP cache

["Sergey Shelukhin (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-20380?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16591009#comment-16591009 ] 

Sergey Shelukhin edited comment on HIVE-20380 at 8/24/18 12:59 AM:
-------------------------------------------------------------------

After trying various approaches I think since this will anyway involve memory copying and interleaving buffers, what needs to happen instead is that we need to decrease allocation size after decompression. Which won't move data, either. However, for small cache, wide table case, where the entire cache can become locked, it's not helpful to replace the fully locked cache of 128Kb buffers with 4Kb of data each with 4Kb buffers sitting in cache every 128Kb. So, we'd have to move data. We will not have multiple CBs per the Java buffer object, but merely change allocations so small CBs don't use large cache buffers.

If we do this shrinking before putting data into cache, then unlike regular cache defragmentation, which is complex, we have a set of already locked buffers that are also invisible to anyone else, so we can trivially consolidate within all the buffers allocated by a read, that noone can touch in any way, and free up some large buffers completely and also some parts of the smaller buffers (i.e. if we have 10 ROW_INDEX streams, each with <4Kb of data, but sitting in 128Kb allocs because the ORC file CB size is 128Kb, we can create 10 4Kb buffers within one of those 10, and straight up deallocate 9 remaining 128Kb buffers, plus the 64Kb + 16Kb + 8Kb in the first one). We can also do an extra step (e.g. if we have a single 4Kb-of-data-128Kb-alloc) of allocating a small buffer explicitly (without defragmentation, and with a flag to not split buffers larger than the original for this - no point in creating a 4Kb buffer out of another 128Kb of empty space for this example), and copying there before deallocating the big one. That will be able to pick up all the crumbs created by other consolidations like the one above. Without splitting and retries the allocation can be cheap and safe.
This will be controlled by a waste threshold setting.

Unfortunately this will do slightly less than nothing at all for Hive 2 without the defrag patch. But, if we backport the defrag patch (pending) this will also work for Hive 2.

I may not be able to work on this to completion immediately so just posting a brain dump here for reference. cc [~gopalv]


was (Author: sershe):
After trying various approaches I think since this will anyway involve memory copying and interleaving buffers, what needs to happen instead is that we need to decrease allocation size after decompression. Which won't move data, either. However, for small cache, wide table case, where the entire cache can become locked, it's not helpful to replace the fully locked cache of 128Kb buffers with 4Kb of data each with 4Kb buffers sitting in cache every 128Kb. So, we'd have to move data. We will not have multiple CBs per the Java buffer object, but merely change allocations so small CBs don't use large cache buffers.

If we do this shrinking before putting data into cache, then unlike regular cache defragmentation, which is complex, we have a set of already locked buffers that are also invisible to anyone else, so we can trivially consolidate within all the buffers allocated by a read, that noone can touch in any way, and free up some large buffers completely and also some parts of the smaller buffers (i.e. if we have 10 ROW_INDEX streams, each with <4Kb of data, but sitting in 128Kb allocs because the ORC file CB size is 128Kb, we can create 10 4Kb buffers within one of those 10, and straight up deallocate 9 remaining 128Kb buffers, plus the 64Kb + 16Kb + 8Kb in the first one). We can also do an extra step (e.g. if we have a single 4Kb-of-data-128Kb-alloc) of allocating a small buffer explicitly (without defragmentation, and with a flag to not split buffers larger than the original for this - no point in creating a 4Kb buffer out of another 128Kb of empty space for this example), and copying there before deallocating the big one. That will be able to pick up all the crumbs created by other consolidations like the one above. Without splitting and retries the allocation can be cheap and safe.
This will be controlled by a waste threshold setting.

Unfortunately this will do slightly less than nothing at all for Hive 2 without the defrag patch. But, if we backport the defrag patch (pending) this will also work for Hive 2.

I may not be able to work on this to completion immediately so just posting a brain dump here for reference.

> explore storing multiple CBs in a single cache buffer in LLAP cache
> -------------------------------------------------------------------
>
>                 Key: HIVE-20380
>                 URL: https://issues.apache.org/jira/browse/HIVE-20380
>             Project: Hive
>          Issue Type: Bug
>            Reporter: Sergey Shelukhin
>            Assignee: Sergey Shelukhin
>            Priority: Major
>
> Lately ORC CBs are becoming ridiculously small. First there's the 4Kb minimum (instead of 256Kb), then after we moved metadata cache off-heap, the index streams that are all tiny take up a lot of CBs and waste space. 
> Wasted space can require larger cache and lead to cache OOMs on some workloads.
> Reducing min.alloc solves this problem, but then there's a lot of heap (and probably compute) overhead to track all these buffers. Arguably even the 4Kb min.alloc is too small.
> We should store contiguous CBs in the same buffer; to start, we can do it for ROW_INDEX streams. That probably means reading all ROW_INDEX streams instead of doing projection when we see that they are too small.
> We need to investigate what the pattern is for ORC data blocks. One option is to increase min.alloc and then consolidate multiple 4-8Kb CBs, but only for the same stream. However larger min.alloc will result in wastage for really small streams, so we can also consolidate multiple streams (potentially across columns) if needed. This will result in some priority anomalies but they probably ok.
> Another consideration is making tracking less object oriented, in particular passing around integer indexes instead of objects and storing state in giant arrays somewhere (potentially with some optimizations for less common things), instead of every buffers getting its own object. 
> cc [~gopalv] [~prasanth_j]



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)