21.6.15.20 The ndbinfo diskpagebuffer Table

The diskpagebuffer table provides statistics about disk page buffer usage by NDB Cluster Disk Data tables.

The diskpagebuffer table contains the following columns:

• node_id

  The data node ID

• block_instance

  Block instance

• pages_written

  Number of pages written to disk.

• pages_written_lcp

  Number of pages written by local checkpoints.

• pages_read

  Number of pages read from disk

• log_waits

  Number of page writes waiting for log to be written to disk

• page_requests_direct_return

  Number of requests for pages that were available in buffer

• page_requests_wait_queue

  Number of requests that had to wait for pages to become available in buffer

• page_requests_wait_io

  Number of requests that had to be read from pages on disk (pages were unavailable in buffer)

Notes

You can use this table with NDB Cluster Disk Data tables to determine whether DiskPageBufferMemory is sufficiently large to allow data to be read from the buffer rather from disk; minimizing disk seeks can help improve performance of such tables.

You can determine the proportion of reads from DiskPageBufferMemory to the total number of reads using a query such as this one, which obtains this ratio as a percentage:

SELECT
  node_id,
  100 * page_requests_direct_return /
    (page_requests_direct_return + page_requests_wait_io)
      AS hit_ratio
FROM ndbinfo.diskpagebuffer;

The result from this query should be similar to what is shown here, with one row for each data node in the cluster (in this example, the cluster has 4 data nodes):

+---------+-----------+
| node_id | hit_ratio |
+---------+-----------+
|       5 |   97.6744 |
|       6 |   97.6879 |
|       7 |   98.1776 |
|       8 |   98.1343 |
+---------+-----------+
4 rows in set (0.00 sec)

hit_ratio values approaching 100% indicate that only a very small number of reads are being made from disk rather than from the buffer, which means that Disk Data read performance is approaching an optimum level. If any of these values are less than 95%, this is a strong indicator that the setting for DiskPageBufferMemory needs to be increased in the config.ini file.

Note

A change in DiskPageBufferMemory requires a rolling restart of all of the cluster's data nodes before it takes effect.

block_instance refers to an instance of a kernel block. Together with the block name, this number can be used to look up a given instance in the threadblocks table. Using this information, you can obtain information about disk page buffer metrics relating to individual threads; an example query using LIMIT 1 to limit the output to a single thread is shown here:

mysql> SELECT
     >   node_id, thr_no, block_name, thread_name, pages_written,
     >   pages_written_lcp, pages_read, log_waits,
     >   page_requests_direct_return, page_requests_wait_queue,
     >   page_requests_wait_io
     > FROM ndbinfo.diskpagebuffer
     >   INNER JOIN ndbinfo.threadblocks USING (node_id, block_instance)
     >   INNER JOIN ndbinfo.threads USING (node_id, thr_no)
     > WHERE block_name = 'PGMAN' LIMIT 1\G
*************************** 1. row ***************************
                    node_id: 1
                     thr_no: 1
                 block_name: PGMAN
                thread_name: rep
              pages_written: 0
          pages_written_lcp: 0
                 pages_read: 1
                  log_waits: 0
page_requests_direct_return: 4
   page_requests_wait_queue: 0
      page_requests_wait_io: 1
1 row in set (0.01 sec)