Thursday May 28, 2009

Inniostat - InnoDB IO Statistics

I wrote a small DTrace script to understand InnoDB IO statistics. This script shows statistics about different kinds of Innodb IO requests and how many of them result in actual IO. Sample output is shown below
#./inniostat -h
Usage: inniostat [-h] [-d] [-p pid] [interval]
                 -h : Print this message
                 -p : MySQL PID
                 -d : Dump dtrace script being used

# ./inniostat
 __physical__  ___Innodb___ ____read____     ______write______
   r/s    w/s    r/s    w/s   data    pre    log dblbuf dflush     Time
    24    121     24     50     24      0     50      0      0 16:00:57
    26    130     26     51     26      0     51      0      0 16:00:58
    18    134     18     54     18      0     54      0      0 16:00:59
    25    129     25     51     25      0     51      0      0 16:01:00
    29    116     46     47     17     29     47      0      0 16:01:01
    10    140     10    132     10      0     52      0     80 16:01:02
    29    129     35     53     14     21     53      0      0 16:01:03

r/s Physical (Actual) writes per sec
w/s Physical (Actual) writes per sec
r/s Reads issued by Innodb per sec
w/s Writes issued by Innodb per sec
data Regular Reads issued by Innodb to the datafiles per sec
pre Prefetch Reads issued by Innodb to the datafiles per sec
log Log writes issued by Innodb per sec
dblbuf Double buffer writes per sec
dflush Writes due to flushing of Innodb buffers to disk.

For the above example you can see that writes to the Innodb double buffer (dblbuf) is zero. This is because I had turned it off explicitly. You will also notice that there are more physical writes than what Innodb is requesting. This can be due to a variety of factors (other engines are doing writes, other applications are doing writes, recordsize mismatch between filesystem and mysql, filesystem logging, etc.). You will also notice a few prefetch requests are being issued.

Another example

Consider the sample output below (contributed by Dimitri)

 __physical__  ___Innodb___ ____read____    ______write______
   r/s    w/s    r/s    w/s   data    pre    log dblbuf dflush     Time
     0   1681      0   5259      0      0   2780      0   2479 14:58:13
     0    224      0   6111      0      0   6111      0      0 14:58:18
     0    761      0   4300      0      0   3368      0    932 14:58:23
     0    503      0   4232      0      0   3546      0    686 14:58:28
     0    904      0   4024      0      0   2923      0   1101 14:58:33
     0   1231      0   4046      0      0   2470      0   1575 14:58:38
     0    502      0   4192      0      0   3640      0    552 14:58:43
     0    928      0   4669      0      0   3331      0   1338 14:58:48
You can quickly notice that this is a cached workload (r/s is zero). There is a significant difference between log writes issued by Innodb and physical writes. This is because this particular setup had innodb_flush_log_at_trx_commit set to 2 (i.e flush log writes every second). If want to flush the log after every commit, you need a disk[s] that can handle around 3000 writes per second. You will also notice a lot of Innodb buffers being flushed (dflush). This maybe because modified buffers in the Innodb buffer pool are being flushed, and/or due to caching in the filesystem instead of in the InnoDB buffer pool.

But I can get some these statistics via "show status"

Some (but not all) of these statistics are available via "show status" command. Be sure to understand what they are measuring before using them. Few of them can be little puzzling. For ex, Innodb_buffer_pool_read_ahead_seq variable counts the number of times a sequential read ahead was triggered. It does not tell you, how many pages were fetched as a result of the read ahead. Similarly, Innodb_log_write_requests tells you how many times a write happened to the log, but its is not really useful as this is incremented when someone writes to the in-memory log block, it does not tell you when the log was written to disk.


Innodb issues reads to either the data files or log files (during recovery). This script does not differentiate between them. Writes are either to the data files, log files, or double write buffer. If you view the source of the script (or use the -d option), you will see that I am interposing on the fil_io function and interpreting its arguments to differentiate IO types. I am using the io:::start probe to figure out the physical IO. Note that this script tracks Innodb IO requests. The actual IO may be carried out by a Innodb background thread some time later. This script also makes assumptions regarding several Innodb constants and hence may not be accurate for future versions.


You can download it here. Since it uses DTrace, it only works on Solaris, MacOS and FreeBSD. Please feel free to use it and let me know of any feedback/comments.

Tuesday May 26, 2009

MySQL Innodb ZFS Best Practices

One of the cool things about talking about MySQL performance with ZFS is that there is not much tuning to be done :-) Tuning with ZFS is considered evil, but a necessity at times. In this blog I will describe some of the tunings that you can apply to get better performance with ZFS as well as point out performance bugs which when fixed will nullify the need for some of these tunings.

For the impatient, here is the summary. See below for the reasoning behind these recommendations and some gotchas.

  1. Match ZFS recordsize with Innodb page size (16KB for Innodb Datafiles, and 128KB for Innodb log files).
  2. If you have a write heavy workload, use a Seperate ZFS Intent Log.
  3. If your database working set size does not fit in memory, you can get a big boost by using a SSD as L2ARC.
  4. While using storage devices with battery backed caches or while comparing ZFS with other filesystems, turn off the cache flush.
  5. Prefer to cache within MySQL/Innodb over the ZFS Adaptive replacement cache (ARC).
  6. Disable ZFS prefetch.
  7. Disable Innodb double write buffer.

Lets look at all of them in detail.

What Match ZFS recordsize with Innodb page size (16KB for Datafiles, and 128KB for Innodb log files).
How zfs set recordsize=16k tank/db

The biggest boost in performance can be obtained by matching the ZFS record size with the size of the IO. Since a Innodb Page is 16KB in size, most read IO is of size 16KB (except for some prefetch IO which can get coalesced). The default recordsize for ZFS is 128KB. The mismatch between the read size and the ZFS recordsize can result in severely inflated IO. If you issue a 16KB read and the data is not already there in the ARC, you have to read 128KB of data to get it. ZFS cannot do a small read because the checksum is calculated for the whole block and you have to read it all to verify data integrity. The other reason to match the IO size and the ZFS recordsize is the read-modify-write penalty. With a ZFS recordsize of 128KB, When Innodb modifies a page, if the zfs record is not already in memory, it needs to be read in from the disk and modified before writing to disk. This increases the IO latency significantly. Luckily matching the ZFS recordsize with the IO size removes all the problems mentioned above.

For Innodb log file, the writes are usually sequential and varying in size. By using ZFS recordsize of 128KB you amortize the cost of read-modify-write.


You need to set the recordsize before creating the database files. If you have already created the files, you need to copy the files to get the new recordsize. You can use the stat(2) command to check the recordsize (look for IO Block:)

What If you have a write heavy workload, use a seperate intent log (slog).
How zpool add log c4t0d0 c4t1d0

Write latency is extremely critical for many MySQL workloads. Typically, a query will read some data, do some calculations, update some data and then commit the transaction. To commit, the Innodb log has to be updated. Many transactions can be committing at the same time. It is very important that this "wait" for commit be fast. Luckily in ZFS, synchronous writes can be accelerated up by using the Seperate Intent Log. In our tests with Sysbench read-write, we have seen around 10-20% improvement with the slog.


What L2ARC (or Level 2 ARC)
How zpool add cache c4t0d0

If your database does not fit in memory, every time you miss the database cache, you have to read a block from disk. This cost is quite high with regular disks. You can minimize the database cache miss latency by using a (or multiple) SSDs as a level-2 cache or L2ARC. Depending on your database working set size, memory and L2ARC size you may see several orders of magnitude improvement in performance.


What When it is safe, turn off ZFS cache flush

The ZFS Evil tuning guide has more information about setting this tunable. Refer to it for the best way to achieve this.


ZFS is designed to work reliably with disks with caches. Everytime it needs data to be stored persistantly on disk, it issues a cache flush command to the disk. Disks with a battery backed caches need not do anything (i.e the cache flush command is a nop). Many storage devices interpret this correctly and do the right thing when they receive a cache flush command. However, there are still a few storage systems which do not interpret the cache flush command correctly. For such storage systems, preventing ZFS from sending the cache flush command results in a big reduction in IO latency. In our tests with Sysbench read-write test we saw a 30% improvement in performance.

  • Setting this tunable on a system without a battery backed cache can cause inconsistencies in case of a crash.
  • When comparing ZFS with filesystems that blindly enable the write cache, be sure to set this to get a fair comparison.

What Prefer to cache within MySQL/Innodb over the ARC.
How Via my.cnf and by limiting the ARC size

You have multiple levels of caching when you are using MySQL/Innodb with ZFS. Innodb has its own buffer pool and ZFS has the ARC. Both of them make independent decisions on what to cache and what to flush. It is possible for both of them to cache the same data. By caching inside Innodb, you get a much shorter (and faster) code path to the data. Moreover, when the Innodb buffer cache is full, a miss in the Innodb buffer cache can lead to flushing of a dirty buffer, even if the data was cached in the ARC. This leads to unnecessary writes. Even though the ARC dynamically shrinks and expands relative to memory pressure, it is more efficient to just limit it.In our tests, we have found that it is better (7-200%) to cache inside Innodb rather than ZFS.


The ARC can be tuned to cache everything, just metadata or nothing on a per filesystem basis. See below for tuning advise about this.

What Disable ZFS Prefetch.
How In /etc/system: set zfs:zfs_prefetch_disable = 1

Most filesystems implement some kind of prefetch. ZFS prefetch detects linear (increasing and decreasing), strided, multiblock strided IO streams and issues prefetch IO when it will help performance. These prefetch IO have a lower priority than regular reads and are generally very beneficial. ZFS also has a lower level prefetch (commonly called vdev prefetch) to help with spatial locality of data.

In Innodb, rows are stored in order of primary index. Innodb issues two kinds of prefetch requests; one is triggered while accessing sequential pages and other is triggered via random access in an extent. While issuing prefetch IO, Innodb assumes that file is laid out in the order of the primary key. This is not true for ZFS. We are yet to investigate the impact of Innodb prefetch.

It is well known that OLTP workloads access data in a random order and hence do not benefit from prefetch. Thus we recommend that you turn off ZFS prefetch.

  • If you have changed the primary cache caching strategy to just cache metadata, you will not trigger file level prefetch.
  • If you have set recordsize to 16k, you will not trigger the lower level prefetch.

What Disable Innodb Double write buffer.
How skip-innodb_doublewrite in my.cnf

Innodb uses a double write buffer for safely updating pages in a tablespace. Innodb first writes the changes to the double write buffer before updating the data page. This is to prevent partial writes. Since ZFS does not allow partial writes, you can safely turn off the double write buffer. In our tests with Sysbench read-write, we say a 5% improvement in performance.





« May 2009 »