🚀 go-pugleaf

On 22 Oct 1995, Brian Tao wrote:

> In article <Pine.SGI.3.91.951019221314.19568B-100000@chaos.ph.utexas.edu>, Alex Predtechensky  <aap@chaos.ph.utexas.edu> wrote:
> >
> > Second, do not use in your tests the 8k block size. There is a linear
> > degradation in performance (to zero) if this size is less than 16k -
> > by some reason, drives do not like short blocks (cache problem?). The
> > maximum performance happens at about 60+-2 k per block, (not 64 or
> > more by unknown reason). 
> 
>     Don't decent drivers perform write clustering to avoid sending a
> whole bunch of small write requests to the drive?
> -- 
> Brian Tao (BT300, taob@io.org)

Yes, they do. But at the expence of additional CPU time. If the access to 
hardware goes without normal DMA (bus-master DMA in PC-terminology),  as 
not so long ago in most of PCs, the overall performance drops by factor 
of two. Even in Linux this write clustering is set to 32 KBytes, and the 
performance is not the maximum one. (This was measured indirectly, just by 
comparing performance tests under DOS with variable block sizes, and  
Linux.


-- Alex