Another thing to consider is how many disk block pointers you have room for in the inode, and when you'll need to add additional indirect blocks. Ref: http://files.gpfsug.org/presentations/2016/south-bank/ D2_P2_A_spectrum_scale_metadata_dark_V2a.pdf
If I understand that presentation correctly.. a block pointer is 12 bytes, so a 1 KB inode can hold approximately 70 block pointers -- ie. any file larger than 70 blocks will need additional indirect blocks to hold the addresses, requiring minimum one additional sub-block. With a 1 MB blocksize filesystem (data and metadata), files between 75 MB and 3 GB will need 1x inode and 1x indirect block = 1 KB + 32 KB metadata space. If your files are smaller than 75 MB, the address pointers fit in the 1K inode. If you have 4 KB inode, files smaller than 340 MB will fit all pointers in the inode. -jf On Wed, Mar 15, 2017 at 2:26 PM, Lukas Hejtmanek <[email protected]> wrote: > On Wed, Mar 15, 2017 at 01:12:44PM +0000, GORECKI, DIETER wrote: > > One other thing to consider is the storage of data inside the inode > itself for very small files. GPFS has the ability to use the remaining > [kilo]bytes of the inode to store the data of the file whenever the file is > small enough to fit in. > > > > Anyone correct me if I am wrong, but with 4k inodes, you can store up to > (4096-128 header) 3968 bytes of data. (without ILM) > > > > So regarding the size of the files you intend to store into your > filesystem, it might be very interesting to take advantage of the > performance of your SSD's to store small files. > > I agree it would, however, I have 30 % free capacity on SSDs only right > now (with > some snapshots and ~70M files). So I'm afraid I have to either change the > rotational disks to hold metadata as well or do not create more > files/snapshots or decrease the inode size. I cannot add more SSDs as I do > not > have free disk slots in HW. > > -- > Lukáš Hejtmánek > _______________________________________________ > gpfsug-discuss mailing list > gpfsug-discuss at spectrumscale.org > http://gpfsug.org/mailman/listinfo/gpfsug-discuss >
_______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
