[Gluster-devel] glusterfs and memcached

[Снежана Бекова]

Thanks for answer.
About using performance translators like io-cache, read-ahead and  
write-behind only in
glusterfs I'm having problem about achieving better performance. I'm  
running glusterfs
2.0.2-1 debian with patched fuse (fuse-2.7.4glfs11) on 3 test machines  
with AFR with client
side replication (2 servers and 1 client). Here is results from  
bonnie++ tests running on the
client glusterfs.

Without performance translators:									 ------Sequential  
Output------ --Sequential Input- --Random-
		 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Time	Size 	 K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec %CP
18m28s	50M	 2012  4      1168    0     730     0     8020   22    7960  
   0       531.7 0
		 ------Sequential Create------ --------Random Create--------
files:max:min                       -Create-- --Read--- -Delete--  
-Create-- --Read--- -Delete--
10:15000:100/100              /sec %CP  /sec %CP  /sec %CP  /sec %CP   
/sec %CP  /sec %CP
		 37    0        135  0        92	  0        44    0        82    0	   
    87   0

With io-threads translator on glusterfs servers:
		 ------Sequential Output------ --Sequential Input- --Random-
		 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Time	Size 	 K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec %CP
32m35s	50M	 1415  2       1024  0      670    0       8112  21    8333  
   0       488.7 0
		 ------Sequential Create------ --------Random Create--------
files:max:min                       -Create-- --Read--- -Delete--  
-Create-- --Read--- -Delete--
10:15000:100/100              /sec %CP  /sec %CP  /sec %CP  /sec %CP   
/sec %CP  /sec %CP
         	                      16     0       68     0       94	   0   
     16     0       58     0	    95     0

With io-cache and write-behind translators on glusterfs client:
		 ------Sequential Output------ --Sequential Input- --Random-
		 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Time	Size 	 K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec %CP
37m55s	50M	 2431  5      2517   0      403    0       8073  22    +++   
  +++  72.2  0
		 ------Sequential Create------ --------Random Create--------
files:max:min                       -Create-- --Read--- -Delete--  
-Create-- --Read--- -Delete--
10:15000:100/100              /sec %CP  /sec %CP  /sec %CP  /sec %CP   
/sec %CP  /sec %CP
         	                       13    0       125   0       91     0   
      13    0        82    0	    78     0

With io-cache translator on glusterfs client:
		 ------Sequential Output------ --Sequential Input- --Random-
		 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Time	Size 	 K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec %CP
20m28s	50M	 1797  3      1177   0      389     0      8081  22    +++   
  +++  75.4  0
		 ------Sequential Create------ --------Random Create--------
files:max:min                       -Create-- --Read--- -Delete--  
-Create-- --Read--- -Delete--
10:15000:100/100              /sec %CP  /sec %CP  /sec %CP  /sec %CP   
/sec %CP  /sec %CP
		 40    0        134   0      104    0      42    0        80     0	   
   92     0


The best result is using glusterfs without performance translators  
especially small files. How to achieve best performance with small  
files?

Thanks,
Snezhana

Цитат от Shehjar Tikoo <shehjart at gluster.com>:

> Снежана Бекова wrote:
>> Hello, I found this very interesting publucation about achieving high
>> performance with memcached and glusterfs:   
>> http://nowlab.cse.ohio-state.edu/publications/conf-papers/2008/noronha-icpp08.pdf
>>
>>
>>
>>
>> Anybody runnig this? Info on how to achive such config with memcached?
>>
>
> Thanks for bringing it up.
> Yes, we're aware of this work but we feel that they've taken a
> long-winded approach to provide caching of file attributes and data by
> using mem-cached. It is an interesting approach but the combination of
> io-cache, read-ahead and write-behind, while being far far simpler in
> design and usage, already handles a large number of
> use cases envisioned by the authors.
>
> Of course, if you're considering using memcached-based approach, we'd
> be interested in hearing about your experience with glusterfs+memcached
> combo.
>
> Regards
> Shehjar
>
>
>> Thanks, Snezhana
>>
>>
>>
>>
>> _______________________________________________ Gluster-devel mailing
>> list Gluster-devel at nongnu.org   
>> http://lists.nongnu.org/mailman/listinfo/gluster-devel
>>