tobert

# enable assertions.  disabling this in production will give a modest
# performance benefit (around 5%).
JVM_OPTS="$JVM_OPTS -ea"

# min and max heap sizes should be set to the same value to avoid
# stop-the-world GC pauses during resize, and so that we can lock the
# heap in memory on startup to prevent any of it from being swapped
# out.
JVM_OPTS="$JVM_OPTS -Xms${MAX_HEAP_SIZE}"
JVM_OPTS="$JVM_OPTS -Xmx${MAX_HEAP_SIZE}"

tobert

Results:
op rate                   : 1981 [simple:1981]
partition rate            : 1981 [simple:1981]
row rate                  : 1981 [simple:1981]
latency mean              : 64.6 [simple:64.6]
latency median            : 40.4 [simple:40.4]
latency 95th percentile   : 150.0 [simple:150.0]
latency 99th percentile   : 185.3 [simple:185.3]
latency 99.9th percentile : 308.0 [simple:308.0]
latency max               : 633.2 [simple:633.2]

tobert

diff --git a/src/java/org/apache/cassandra/io/util/RandomAccessReader.java b/src/java/org/apache/cassandra/io/util/RandomAccessReader.java
index 95877a2..47b4b9b 100644
--- a/src/java/org/apache/cassandra/io/util/RandomAccessReader.java
+++ b/src/java/org/apache/cassandra/io/util/RandomAccessReader.java
@@ -31,6 +31,7 @@ public class RandomAccessReader extends RandomAccessFile implements FileDataInpu

     // default buffer size, 64Kb
     public static final int DEFAULT_BUFFER_SIZE = 65536;
+    public static final int BUFFER_SIZE = getBufferSize();

tobert

Results:
op rate                   : 7741 [READ:7741]
partition rate            : 7741 [READ:7741]
row rate                  : 7741 [READ:7741]
latency mean              : 103.3 [READ:103.3]
latency median            : 79.6 [READ:79.6]
latency 95th percentile   : 217.7 [READ:217.7]
latency 99th percentile   : 347.9 [READ:347.9]
latency 99.9th percentile : 515.7 [READ:515.7]
latency max               : 977.7 [READ:977.7]

tobert

# cassandra-stress user profile=ycsb.yaml
# Just a first pass. Needs some tweaking.

keyspace: ycsb

keyspace_definition: |
  CREATE KEYSPACE ycsb WITH replication = { 'class' : 'SimpleStrategy', 'replication_factor' : 3 };

table: small

tobert

#!/bin/bash
# A quick & dirty script to pin compaction threads to a core without
# any consideration for HT or what kind of compaction any given thread
# is doing.
# atobey@datastax.com 2015-08-02

# match all processes with nice value 4, which is hard-coded in Cassandra
pids=$(ps -eLo nice,tid,args |awk '/^ *4 .*java/{print $2}')

ncores=$(grep ^processor /proc/cpuinfo |wc -l)

tobert

# jmx: metrics and administration interface
#
# add this if you're having trouble connecting:
# JVM_OPTS="$JVM_OPTS -Djava.rmi.server.hostname=<public name>"

myip=$(ip -4 addr show eno1 |awk '/inet /{ gsub(/\/[0-9]+/, ""); print $2 }')
JVM_OPTS="$JVM_OPTS -Djava.rmi.server.hostname=$myip"

#
# see

tobert

atobey@brak ~ $ sudo smartctl -a /dev/sda
smartctl 6.4 2015-06-04 r4109 [x86_64-linux-4.0.1-1-ARCH] (local build)
Copyright (C) 2002-15, Bruce Allen, Christian Franke, www.smartmontools.org

=== START OF INFORMATION SECTION ===
Vendor:               SEAGATE
Product:              ST9500430SS
Revision:             DS04
User Capacity:        500,107,862,016 bytes [500 GB]
Logical block size:   512 bytes

tobert

atobey@brak ~ $ sudo ioping -c 10 -S 4k -D /mnt/sdg
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=1 time=391 us
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=2 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=3 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=4 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=5 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=6 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=7 time=8.62 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=8 time=8.63 ms
4 KiB from /mnt/sdg (xfs /dev/sdg1): request=9 time=8.93 ms

tobert

Ran on a 15 node cluster. 6GB heap with 2GB offheap memtables, G1GC, and jemalloc.

Small is 16 cols. Medium is 30 cols. Large is 60 cols wide. All 3 ran simultaneously from 2 client nodes running cassandra-stress.