预估ceph的迁移数据量

引言

我们在进行 ceph 的 osd 的增加和减少的维护的时候，会碰到迁移数据，但是我们平时会怎么去回答关于迁移数据量的问题，一般来说，都是说很多，或者说根据环境来看，有没有精确的一个说法，到底要迁移多少数据?这个我以前也有思考过这个问题，当时想是对比前后的pg的分布，然后进行计算，正好在翻一些资料的时候，看到有alram写的一篇博客，alram是Inktank的程序员，也就是sage所在的公司，程序是一个python脚本，本篇会分析下这个对比的思路，以及运行效果

计算迁移量只需要一个修改后的crushmap就可以了，这个是离线计算的，所以不会对集群有什么影响

运行效果

准备修改后的crushmap

获取当前crushmap

ceph osd getcrushmap -o crushmap

解码crushmap

crushtool -d crushmap -o crushmap.txt

修改crushmap.txt
这个根据自己需要，修改成自己想修改成的crushmap即可，可以是增加，也可以是删除

减少节点的计算

假如删除一个osd.5 我们需要迁移多少数据
将crushmap里面的osd.5的weight改成0

crushtool -c crushmap.txt -o crushmapnew

运行计算脚本

[root@lab8106 ceph]# python jisuan.py --crushmap-file crushmapnew
POOL                 REMAPPED OSDs        BYTES REBALANCE      OBJECTS REBALANCE   
rbd                  59                   6157238296           1469                
data                 54                   5918162968           1412                
metadata             53                   5825888280           1390

可以看到迁移的数据量
REMAPPED OSDs 下面就是有多少份的PG数据需要迁移，这里面计算的方式是比较前后的分布

[1,2] - > [1,2] 迁移0个

[1,2] - > [4,2] 迁移1个

[1,2] - > [4,3] 迁移2个

上面的统计的是这样的个数，所以不太好说是PG或者是OSD，可以理解为PG内数据的份数，因为单个PG可能需要迁移一份，也有可能迁移两份，或者多份

增加节点的计算

如果增加一个osd.6 我们需要迁移多少数据
直接运行脚本

[root@lab8106 ceph]# python jisuan.py --crushmap-file crushmapnew
POOL                 REMAPPED OSDs        BYTES REBALANCE      OBJECTS REBALANCE   
rbd                  0                    0                    0                   
data                 0                    0                    0                   
metadata             0                    0                    0 

可以看到没有输出，这个是因为计算的脚本里面有个地方报错了，ceph内部有个限制，在将crushmap import进osdmap的时候，ceph会验证osdmap里面的osd个数和crushmap里面的osd个数是不是相同
所以这个地方需要多做一步，将osd的个数设置成跟预估的一致，这个是唯一对现有集群做的修改操作，

[root@lab8106 ceph]# ceph osd setmaxosd 7
set new max_osd = 7

然后再次运行就可以了

[root@lab8106 ceph]# python jisuan.py --crushmap-file crushmapnew
POOL                 REMAPPED OSDs        BYTES REBALANCE      OBJECTS REBALANCE   
rbd                  31                   3590324224           856                 
data                 34                   3372220416           804                 
metadata             41                   4492099584           1071                

上面就是运行的效果，下面我们对内部的逻辑进行分析

代码和代码分析

代码

#!/usr/bin/env python

import ast
import json
import os
import subprocess
import argparse
import sys

FNULL = open(os.devnull, 'w')

# assume the osdmap test output
# is the same lenght and order...
# if add support for PG increase
# that's gonna break
def diff_output(original, new, pools):
    number_of_osd_remap = 0
    osd_data_movement = 0

    results = {}

    pg_data, pg_objects = get_pg_info()

    for i in range(len(original)):
        orig_i = original[i]
        new_i = new[i]

        if orig_i[0].isdigit():
            pg_id = orig_i.split('	')[0]
            pool_id = pg_id[0]
            pool_name = pools[pool_id]['pool_name']

            if not pool_name in results:
                results[pool_name] = {}
                results[pool_name]['osd_remap_counter'] = 0
                results[pool_name]['osd_bytes_movement'] = 0
                results[pool_name]['osd_objects_movement'] = 0

            original_mappings = ast.literal_eval(orig_i.split('	')[1])
            new_mappings = ast.literal_eval(new_i.split('	')[1])
            intersection = list(set(original_mappings).intersection(set(new_mappings)))

            osd_movement_for_this_pg = int(pools[pool_id]['pool_size']) - len(intersection)
            osd_data_movement_for_this_pg = int(osd_movement_for_this_pg) * int(pg_data[pg_id])
            osd_object_movement_for_this_pg = int(osd_movement_for_this_pg) * int(pg_objects[pg_id])

            results[pool_name]['osd_remap_counter'] += osd_movement_for_this_pg
            results[pool_name]['osd_bytes_movement'] += int(osd_data_movement_for_this_pg)
            results[pool_name]['osd_objects_movement'] += int(osd_object_movement_for_this_pg)

        elif orig_i.startswith('#osd'):
            break

    return results

def get_pools_info(osdmap_path):
    pools = {}
    args = ['osdmaptool', '--print', osdmap_path]
    osdmap_out = subprocess.check_output(args, stderr=FNULL).split('
')
    for line in osdmap_out:
        if line.startswith('pool'):
            pool_id = line.split()[1]
            pool_size = line.split()[5]
            pool_name = line.split()[2].replace("'","")
            pools[pool_id] = {}
            pools[pool_id]['pool_size'] = pool_size
            pools[pool_id]['pool_name'] = pool_name
        elif line.startswith('max_osd'):
            break

    return pools

def get_osd_map(osdmap_path):
    args = ['sudo', 'ceph', 'osd', 'getmap', '-o', osdmap_path]
    subprocess.call(args, stdout=FNULL, stderr=subprocess.STDOUT)

def get_pg_info():
    pg_data = {}
    pg_objects = {}
    args = ['sudo', 'ceph', 'pg', 'dump']
    pgmap = subprocess.check_output(args, stderr=FNULL).split('
')

    for line in pgmap:
        if line[0].isdigit():
            pg_id = line.split('	')[0]
            pg_bytes = line.split('	')[6]
            pg_obj = line.split('	')[1]
            pg_data[pg_id] = pg_bytes
            pg_objects[pg_id] = pg_obj
        elif line.startswith('pool'):
            break

    return pg_data, pg_objects

def osdmaptool_test_map_pgs_dump(original_osdmap_path, crushmap):
    new_osdmap_path = original_osdmap_path + '.new'
    get_osd_map(original_osdmap_path)
    args = ['osdmaptool', '--test-map-pgs-dump', original_osdmap_path]
    original_osdmaptool_output = subprocess.check_output(args, stderr=FNULL).split('
')

    args = ['cp', original_osdmap_path, new_osdmap_path]
    subprocess.call(args, stdout=FNULL, stderr=subprocess.STDOUT)
    args = ['osdmaptool', '--import-crush', crushmap, new_osdmap_path]
    subprocess.call(args, stdout=FNULL, stderr=subprocess.STDOUT)
    args = ['osdmaptool', '--test-map-pgs-dump', new_osdmap_path]
    new_osdmaptool_output = subprocess.check_output(args, stderr=FNULL).split('
')

    pools = get_pools_info(original_osdmap_path)
    results = diff_output(original_osdmaptool_output, new_osdmaptool_output, pools)

    return results


def dump_plain_output(results):
    sys.stdout.write("%-20s %-20s %-20s %-20s
" % ("POOL", "REMAPPED OSDs", "BYTES REBALANCE", "OBJECTS REBALANCE"))

    for pool in results:
        sys.stdout.write("%-20s %-20s %-20s %-20s
" % (
            pool,
            results[pool]['osd_remap_counter'],
            results[pool]['osd_bytes_movement'],
            results[pool]['osd_objects_movement']
            ))

def cleanup(osdmap):
    FNULL.close()
    new_osdmap = osdmap + '.new'
    os.remove(new_osdmap)

def parse_args():
    parser = argparse.ArgumentParser(description='Ceph CRUSH change data movement calculator.')

    parser.add_argument(
        '--osdmap-file',
        help="Where to save the original osdmap. Temp one will be <location>.new. Default: /tmp/osdmap",
        default="/tmp/osdmap",
        dest="osdmap_path"
        )
    parser.add_argument(
        '--crushmap-file',
        help="CRUSHmap to run the movement test against.",
        required=True,
        dest="new_crushmap"
        )

    parser.add_argument(
        '--format',
        help="Output format. Default: plain",
        choices=['json', 'plain'],
        dest="format",
        default="plain"
        )

    args = parser.parse_args()
    return args

if __name__ == '__main__':
    ctx = parse_args()

    results = osdmaptool_test_map_pgs_dump(ctx.osdmap_path, ctx.new_crushmap)
    cleanup(ctx.osdmap_path)

    if ctx.format == 'json':
        print json.dumps(results)
    elif ctx.format == 'plain':
        dump_plain_output(results)

直接放在这里方便拷贝，也可以去原作者的gist里面去获取

主要代码分析

首先获取osdmap

ceph osd getmap -o /tmp/osdmap

获取原始pg分布

使用osdmaptool  --test-map-pgs-dump /tmp/osdmap

获取新的crushmap

这个是自己编辑成需要的crushmap

将新的crushmap注入到osdmap里面得到新的osdmap

osdmaptool --import-crush  crushmap  /tmp/new_osdmap_path

根据新的osdmap进行计算新的分布

osdmaptool  --test-map-pgs-dump /tmp/new_osdmap_path

然后比较两个输入进行对比得到结果

相关链接

Calculate data migration when changing the CRUSHmap
alram/crush_data_movement_calculator.py

变更记录

Why	Who	When
创建	武汉-运维-磨渣	2017-02-08