[ https://issues.apache.org/jira/browse/HBASE-5987?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13273000#comment-13273000 ]
Todd Lipcon commented on HBASE-5987: ------------------------------------ Nice stuff Liyin. I was looking at scan performance a bit last week as well and came to similar conclusions that the reseeks were pretty expensive for this reason. Another thing I noticed is that our CPU cache behavior is pretty bad when the individual KVs are large. When I profiled L2 cache misses in oprofile, I saw a bunch on the call to read the memstoreTS -- assumedly because it fell on a different cache line than the rest of the KV. In my case, the KVs were just over 128 bytes (2 cache lines), including their header fields, lengths etc. So the access pattern looked like: - hit cacheline 0 for kv0 header - hit cacheline 2 for kv0 memstoreTS - hit cacheline 0 repeatedly to do KV comparison - hit cacheline 2 for kv1's header - hit cacheline 4 for kv1's memstoreTS - hit cacheline 2 for kv1 data comparison etc. For whatever reason, my CPU wasn't quite smart enough to kick prefetching in on this access pattern. I tried recompiling JDK7 with the Unsafe.prefetchRead intrinsic, but couldn't get any noticeable improvement with it. So I think for better performance, we need some better in-memory layout for the HFile blocks, so we can get O(lg n) reseeks instead of O(n), for example. Have you seen the same? > HFileBlockIndex improvement > --------------------------- > > Key: HBASE-5987 > URL: https://issues.apache.org/jira/browse/HBASE-5987 > Project: HBase > Issue Type: Improvement > Reporter: Liyin Tang > Assignee: Liyin Tang > > Recently we find out a performance problem that it is quite slow when > multiple requests are reading the same block of data or index. > From the profiling, one of the causes is the IdLock contention which has been > addressed in HBASE-5898. > Another issue is that the HFileScanner will keep asking the HFileBlockIndex > about the data block location for each target key value during the scan > process(reSeekTo), even though the target key value has already been in the > current data block. This issue will cause certain index block very HOT, > especially when it is a sequential scan. > To solve this issue, we propose the following solutions: > First, we propose to lookahead for one more block index so that the > HFileScanner would know the start key value of next data block. So if the > target key value for the scan(reSeekTo) is "smaller" than that start kv of > next data block, it means the target key value has a very high possibility in > the current data block (if not in current data block, then the start kv of > next data block should be returned. +Indexing on the start key has some > defects here+) and it shall NOT query the HFileBlockIndex in this case. On > the contrary, if the target key value is "bigger", then it shall query the > HFileBlockIndex. This improvement shall help to reduce the hotness of > HFileBlockIndex and avoid some unnecessary IdLock Contention or Index Block > Cache lookup. > Secondary, we propose to push this idea a little further that the > HFileBlockIndex shall index on the last key value of each data block instead > of indexing on the start key value. The motivation is to solve the HBASE-4443 > issue (avoid seeking to "previous" block when key you are interested in is > the first one of a block) as well as +the defects mentioned above+. > For example, if the target key value is "smaller" than the start key value of > the data block N. There is no way for sure the target key value is in the > data block N or N-1. So it has to seek from data block N-1. However, if the > block index is based on the last key value for each data block and the target > key value is beween the last key value of data block N-1 and data block N, > then the target key value is supposed be data block N for sure. > As long as HBase only supports the forward scan, the last key value makes > more sense to be indexed on than the start key value. > Thanks Kannan and Mikhail for the insightful discussions and suggestions. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira