2009-08-27

• [計畫] Hadoop 叢集維護
  • [狀況] 發現 hadoop104, hadoop106 kernel panic
• [計畫] Hadoop 叢集維護
  • [狀況] 發現 /etc/hadoop/conf/hadoop-site.xml 中 dfs.replication 數值為 1 也就是沒做備份
  • [解法]
    1. 修改 dfs.replication 數值為 3
    2. 重新執行 hadoop-namenode 與 hadoop-datanode
    3. 使用 hadoop fs -setrep 設定目前為 1 的 /user 目錄所有檔案

       root@hadoop:~# su -s /bin/sh hadoop -c "hadoop fs -setrep -R 3 /user"
       

    4. 使用 hadoop balancer 嘗試資料的 replication 機制是否會被執行

       root@hadoop:~# su -s /bin/sh hadoop -c "hadoop balancer"
       

    5. 使用 hadoop fsck 嘗試資料的 replication 機制是否會被執行

       root@hadoop:~# su -s /bin/sh hadoop -c "hadoop fsck / -racks"
       

       ### 會有訊息顯示目前的 replication 數目不夠
       /user/waue/input/1.txt:  Under replicated blk_-682447276956362627_16045. Target Replicas is 3 but found 1 replica(s).
       ### 自從誤刪 hadoop113 硬碟資料後，HDFS 狀態都是 CORRUPT，看樣子要請大家重新上傳看看了
       Status: CORRUPT
        Total size:    2514937876121 B
        Total dirs:    2800
        Total files:   14972
        Total blocks (validated):      51686 (avg. block size 48658009 B)
         ********************************
         CORRUPT FILES:        1921
         MISSING BLOCKS:       4972
         MISSING SIZE:         232737717270 B
         CORRUPT BLOCKS:       4972
         ********************************
        Minimally replicated blocks:   46714 (90.38037 %)
        Over-replicated blocks:        3 (0.00580428 %)
        Under-replicated blocks:       45388 (87.81488 %)
        Mis-replicated blocks:         0 (0.0 %)
        Default replication factor:    3
        Average block replication:     1.0597067
        Corrupt blocks:                4972
        Missing replicas:              89401 (163.2239 %)
        Number of data-nodes:          17
        Number of racks:               1
       
       The filesystem under path '/' is CORRUPT
       

    6. 跑過 fsck 再跑 balancer 就會看到正在嘗試做資料的 replication 副本動作

       root@hadoop:~# su -s /bin/sh hadoop -c "hadoop balancer"
       
       09/08/27 06:42:34 INFO dfs.Balancer: Need to move 12.21 GB bytes to make the cluster balanced.
       09/08/27 06:42:34 INFO dfs.Balancer: Decided to move 10 GB bytes from 192.168.1.10:50010 to 192.168.1.17:50010
       09/08/27 06:42:34 INFO dfs.Balancer: Decided to move 10 GB bytes from 192.168.1.1:50010 to 192.168.1.19:50010
       09/08/27 06:42:34 INFO dfs.Balancer: Decided to move 10 GB bytes from 192.168.1.11:50010 to 192.168.1.16:50010
       09/08/27 06:42:34 INFO dfs.Balancer: Decided to move 10 GB bytes from 192.168.1.5:50010 to 192.168.1.14:50010
       09/08/27 06:42:34 INFO dfs.Balancer: Decided to move 10 GB bytes from 192.168.1.9:50010 to 192.168.1.18:50010
       09/08/27 06:42:34 INFO dfs.Balancer: Will move 50 GBbytes in this iteration
       2009/8/27 上午 06:42:34             0                 0 KB            12.21 GB              50 GB
       
       09/08/27 07:29:00 INFO dfs.Balancer: Decided to move block -5073029869990347015 with a length of 64 MB bytes from 192.168.1.2:50010 to 192.168.1.15:50010 using proxy source 192.168.1.2:50010
       09/08/27 07:29:00 INFO dfs.Balancer: Starting moving -5073029869990347015 from 192.168.1.2:50010 to 192.168.1.15:50010
       

  • [發現] Hadoop 對於 HDFS /var/lib/hadoop/cache 目錄裡的檔案還真是保護到極致了...設定了 10 個 replication 副本

    -rw-r--r--  10 hadoop002 supergroup     108739 2009-08-24 22:55 /var/lib/hadoop/cache/hadoop/mapred/system/job_200908242228_0009/job.jar
    

• [計畫] Hadoop 叢集維護
  • [狀況] 發現把副本設為 3 會造成空間不足，此外因為檔案數多，但都是小檔案，或許不該把 block size 設為 64Mbytes (67108864)
  • [解法]
    1. 把副本設為 2 並該把 block size 設為 4kbytes (4096)
    2. '由於 Block Size 只有檔案產生的時候才會生效,因此目前看起來除了跑 fsck 外，還要跑 balancer，有搬動過的才有機會變小。其次就是用 hadoop fs -cp /user /user2 比較有機會全面性的造成影響。

       09/08/27 14:37:31 INFO dfs.Balancer: Decided to move block -2997460447318763427 with a length of 4 KB bytes from 192.168.1.18:50010 to 192.168.1.2:50010 using proxy source 192.168.1.18:50010
       09/08/27 14:37:31 INFO dfs.Balancer: Starting moving -2997460447318763427 from 192.168.1.18:50010 to 192.168.1.2:50010
       09/08/27 14:37:31 INFO dfs.Balancer: Decided to move block -2997680912701056691 with a length of 4 KB bytes from 192.168.1.18:50010 to 192.168.1.2:50010 using proxy source 192.168.1.18:50010
       09/08/27 14:37:31 INFO dfs.Balancer: Starting moving -2997680912701056691 from 192.168.1.18:50010 to 192.168.1.2:50010
       09/08/27 14:37:31 INFO dfs.Balancer: Decided to move block 248102854142724550 with a length of 13.36 KB bytes from 192.168.1.6:50010 to 192.168.1.5:50010 using proxy source 192.168.1.6:50010
       

  • [統計] 目前平均 file size 大概都落在 1KB 以下。
  • 
  • [發現] 根據 Hadoop 論壇的討論，發現其實 block size 是由使用者自己(Hadoop Client)決定的。只要在上傳或複製時，用 -D 指定 dfs.block.size 屬性，就可以自訂 block size。底下的範例：
    • /user/jazz/input/hadoop-default.xml - 修改前產生的檔案(/etc/hadoop/conf/hadoop-site.xml 的 dfs.block.size = 67108864 , i.e. 64MB)
    • /user/jazz/input/hadoop-default.xml.new - 修改後複製產生的檔案(/etc/hadoop/conf/hadoop-site.xml 的 dfs.block.size = 4096 , i.e. 4KB)
    • /user/jazz/input/hadoop-default.xml.new - 自訂屬性複製產生的檔案(/etc/hadoop/conf/hadoop-site.xml 的 dfs.block.size = 4096 , i.e. 4KB)

      jazz@hadoop:~$ hadoop fs -cp /user/jazz/input/hadoop-default.xml /user/jazz/input/hadoop-default.xml.new
      jazz@hadoop:~$ hadoop fs -D dfs.block.size=4194304 -cp /user/jazz/input/hadoop-default.xml /user/jazz/input/hadoop-default.xml.4M
      jazz@hadoop:~$ hadoop fs -stat "filesize=%b block_size=%o filename=%n replication=%r" /user/jazz/input/hadoop-default.xml
      filesize=40673 block_size=67108864 filename=hadoop-default.xml replication=1
      jazz@hadoop:~$ hadoop fs -stat "filesize=%b block_size=%o filename=%n replication=%r" /user/jazz/input/hadoop-default.xml.4M
      filesize=40673 block_size=4194304 filename=hadoop-default.xml.4M replication=2
      jazz@hadoop:~$ hadoop fs -stat "filesize=%b block_size=%o filename=%n replication=%r" /user/jazz/input/hadoop-default.xml.new
      filesize=40673 block_size=4096 filename=hadoop-default.xml.new replication=1