Thank you so much again for those so helpful benchmark results.
just a few comments.
On Tue, Apr 01, 2003 at 06:29:50PM -0500, rwhron@earthlink.net wrote:
> File create/delete and VM system latencies in microseconds - smaller is better
> ----------------------------------------------------------------------------
> 0K 0K 1K 1K 4K 4K Mmap Prot Page
> kernel Create Delete Create Delete Create Delete Latency Fault Fault
> ------------------------------ ------- ------- ------- ------- ------- ------- ------ ------
> 2.4.21-pre5-ac3 63.5 15.4 139.7 27.0 146.7 27.0 2630 1.34 5.2
> 2.4.21-pre5-akpm 64.1 14.1 139.5 24.3 144.3 24.3 2541 0.99 5.4
> 2.4.21-pre5 64.8 13.5 134.8 24.8 140.5 25.0 2593 0.84 5.1
> 2.5.66-mjb1 72.4 14.4 154.3 27.8 158.6 27.8 3621 0.66 8.8
> 2.5.66 78.2 15.1 161.0 27.9 166.9 27.9 3887 0.78 8.6
> 2.5.66-ac1 83.4 16.7 166.9 34.3 170.8 34.3 3942 1.25 15.2
> 2.5.66-mm1 89.3 17.0 182.3 35.4 188.0 35.4 4413 0.91 9.5
> 2.4.21-pre5aa2 89.7 14.2 165.3 27.3 174.3 27.3 2480 1.05 5.9
> 2.4.21-pre5-jam1 91.4 14.0 167.7 27.9 172.5 27.9 2508 0.84 6.0
> 2.2.23 141.3 21.7 207.5 27.5 215.4 27.4 64369 0.89 1246.0
the reason my tree is slower in create is intentional: I drop the
negative dentries after unlink to better preserve the working set, and
to release IMHO worthless cache in smart way. I don't think it's common
to unlink and open immediatly back. the other kernels do better here
because they optimize for unlike + open. Of course also in my tree the
first open failure will trigger the reallocation of the negative dentry.
> 2.5.x has lower cpu utilization for sequential block reads.
>
> ---------------------Sequential Output--------------------
> -----Per Char----- ------Block------- -----Rewrite------
> Kernel MB/sec %CPU Eff MB/sec %CPU Eff MB/sec %CPU Eff
> 2.4.21-pre5aa2 3.80 98.0 3.88 22.46 85.7 26.21 9.86 52.0 18.96
> 2.4.21-pre5-jam1 3.77 98.0 3.85 21.99 87.0 25.28 9.81 53.3 18.40
> 2.4.21-pre5 3.76 98.0 3.84 21.43 87.0 24.63 9.62 44.7 21.55
> 2.4.21-pre5-akpm 3.76 98.0 3.84 21.31 87.7 24.31 8.88 34.7 25.61
> 2.4.21-pre5-ac3 3.76 98.0 3.84 21.21 86.7 24.47 9.54 44.3 21.52
> 2.5.66-mjb1 3.69 97.0 3.81 20.96 87.0 24.09 7.97 32.3 24.64
> 2.5.66 3.66 97.0 3.78 20.55 85.0 24.18 7.37 31.7 23.27
> 2.5.66-mm1 3.68 97.0 3.80 20.48 85.0 24.10 7.85 28.0 28.05
> 2.5.66-ac1 3.63 97.0 3.74 20.43 83.7 24.42 7.80 33.3 23.39
> 2.2.23 2.96 73.7 4.02 9.85 59.3 16.60 4.47 87.3 5.11
>
> -----------Sequential Input----------- ------Random-----
> -----Per Char----- ------Block------- ------Seeks------
> Kernel MB/sec %CPU Eff MB/sec %CPU Eff /sec %CPU Eff
> 2.4.21-pre5aa2 3.92 95.7 4.09 21.26 77.3 27.49 140 1.7 8402
> 2.4.21-pre5-jam1 3.92 96.3 4.07 21.93 82.0 26.74 136 1.7 8150
> 2.4.21-pre5 4.01 97.0 4.13 18.30 66.7 27.45 144 1.7 8656
> 2.4.21-pre5-akpm 3.86 94.0 4.11 17.19 52.3 32.85 141 2.0 7038
> 2.4.21-pre5-ac3 4.01 98.0 4.09 18.40 66.3 27.74 138 1.7 8270
> 2.5.66-mjb1 4.00 99.0 4.04 15.00 16.3 91.84 126 3.0 4210
> 2.5.66 3.94 98.3 4.01 14.24 17.0 83.76 137 3.0 4574
> 2.5.66-mm1 3.92 98.0 4.00 14.58 16.7 87.46 154 3.0 5140
> 2.5.66-ac1 3.98 99.0 4.02 14.64 16.3 89.65 134 4.0 3344
> 2.2.23 3.05 96.7 3.15 9.62 77.3 12.43 133 1.0 13329
the improvement for read contigous of my tree versus the others is
nothing here, scsi really shows the difference between my tree and
all others including 2.5. IDE is capable of 64k dma only, so the
difference is not huge in the above results. On scsi a plain bonnie
approches a 100% improvement on some high end hardware as you also can
see in bigbox.html.
thanks,
Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/