Re: Database Usage of Unix FFS

lparsons_at_world.std.com (Lee E Parsons) writes:

>murphyn_at_orca.cig.mot.com (Neal P. Murphy) writes:
>>pete_at_tecc.co.uk (Pete Bentley) writes:
>>
>>>>>>>> lparsons_at_world.std.com (Lee E Parsons) writes:
 

>>>For what you are doing, putting a few big files on a big disk, it is
>>>tempting to suggest that a more traditional, SVR3-type filesystem
>>>might be more efficient...allocate a really small number of inodes,
>>>put all the files as few directories as possible and the SVR3 layout
>>>policy means you can get all the files contiguous if you create them
>>>on another f/s and tar or cpio them to their destination. Comments?
>>
>>For what Lee is doing, would he be better off using raw or block
>>devices, completely bypassing the filesystem? I should think that
>>eliminating a level of processing would enhance performance. With
>>15 slices/partitions on a 1GB disk, he could get 75MB slices, on
>>average.
 

>Raw is probably the best option, one of the things I'm doing is trying
>to is understand the best way of laying down a large database on a
>cooked FS in order to see which is the best.

        My experience is that tuning the filesystem parameters is a big win. One of the most important ones is the maximum number of blocks a file can allocate within a cylinder group before having to move on to the next group. If you can set this to 100% of the cylinder group it helps sequential reads greatly (the default is 25% I think). Also if you can set the rotational delay to 0 (which most modern
hardware can handle), create with the biggest block size you can (8k, 16k, 32k) this is also ye bigge win. Doing this allowed us to get platter speed on Oracle full table scans on a MIPS 6280 with SMD drives (2Mb/s). Those with Seagate Barracudas could undoubtly do better. FFS can be your friend ;^).

        Where accessing raw partitions wins is in the elimination of the buffer copies. This is not so bad when you dont have bazillions of striped drives, but after a certain number of controllers, IO buses et cetera memory bandwidth does become the main issue.

        Stephen

--

	Stephen