[ https://issues.apache.org/jira/browse/CASSANDRA-11940?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15331973#comment-15331973 ]
Eric Evans commented on CASSANDRA-11940: ---------------------------------------- In the plot below, nodes cerium, praseodymium (worst. hostname. evar.), and xenon are under a workload consisting primarily of random reads. At ~11:00, {{disk_access_mode}} is changed from {{auto}} to {{mmap_index_only}} on cerium and praseodymium only. When this happens, disk read throughput drops dramatically, as does the number of major page faults generated by Cassandra (from ~280/s to ~15/s). Over the course of this run, I increased {{file_cache_size_in_mb}} from 512, to 768, 1024, and 2048, with no observable difference to disk read throughput, or the rate of major page faults. These machines are 12-way (w/ hyperthreading), 16G RAM, and rotational disks. They have 4G heaps (though raised to 6G for the {{file_cache_size_in_mb: 2048}} test) !abnormal disk read throughput.png|width=960! {noformat} CREATE TABLE "local_group_wikipedia_T_parsoid_html".data ( "_domain" text, key text, rev int, tid timeuuid, "_del" timeuuid, "content-location" text, "content-sha256" blob, "content-type" text, "latestTid" timeuuid, tags set<text>, value blob, PRIMARY KEY (("_domain", key), rev, tid) ) WITH CLUSTERING ORDER BY (rev DESC, tid DESC) AND bloom_filter_fp_chance = 0.1 AND caching = '{"keys":"ALL", "rows_per_partition":"NONE"}' AND comment = '' AND compaction = {'tombstone_threshold': '0.02', 'unchecked_tombstone_compaction': 'true', 'base_time_seconds': '45', 'class': 'org.apache.cassandra.db.compaction.DateTieredCompactionStrategy'} AND compression = {'chunk_length_kb': '512', 'sstable_compression': 'org.apache.cassandra.io.compress.DeflateCompressor'} AND dclocal_read_repair_chance = 0.0 AND default_time_to_live = 0 AND gc_grace_seconds = 86400 AND max_index_interval = 2048 AND memtable_flush_period_in_ms = 0 AND min_index_interval = 128 AND read_repair_chance = 0.0 AND speculative_retry = '99.0PERCENTILE'; {noformat} > Look into better default file_cache_size for 2.2 > ------------------------------------------------ > > Key: CASSANDRA-11940 > URL: https://issues.apache.org/jira/browse/CASSANDRA-11940 > Project: Cassandra > Issue Type: Improvement > Reporter: T Jake Luciani > Assignee: T Jake Luciani > Fix For: 2.2.x > > Attachments: abnormal disk read throughput.png > > > CASSANDRA-8464 added support for mmapped decompression where in version <= > 2.1 the reads were all decompressed in standard heap buffers. > Since the usage of the file_cache is based solely on the buffer capacity we > should/can make this much larger in 2.2 when the disk access mode is mmap. > The downside of this cache being too small is made worse by 8464 since the > buffers are mmapped/unmapped causing explicit page faults. -- This message was sent by Atlassian JIRA (v6.3.4#6332)