Re: openafs is really faster on linux-2.4. than 2.6

From: Martin MOKREJÅ
Date: Thu Aug 18 2005 - 11:05:32 EST


Hi Con,
thank you for anwers. It seems my main confusion was that values in 'id' and
'wa' columns in vmstat(1) output do not reflect the 2.4 kernel stats well.
The timing shows that the real time is more or less same and we could only argue
if the sys time is significantly higher on 2.6 kernel or not. So, there
ios no problem with the kernel. ;)


on 2.4.31 (read from local xfs disk to another disk with ext2 fs):

procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
r b swpd free buff cache si so bi bo in cs us sy id wa
5 0 0 252932 4172 2637980 0 0 57856 47500 1548 1884 2 27 71 0
1 1 0 136192 4172 2754716 0 0 58368 52804 1599 1885 2 24 74 0
2 1 0 85588 4172 2805052 0 0 61696 51600 1636 1965 5 40 55 0
1 1 0 85376 4172 2804764 0 0 59520 53828 1623 1848 3 45 52 0
3 1 0 85772 4172 2805036 0 0 60288 50112 1609 1875 1 33 66 0
1 1 0 85784 4172 2804364 0 0 58240 50300 1567 1835 2 43 55 0
2 1 0 85468 4172 2805116 0 0 61824 47516 1666 2149 4 47 50 0
1 1 0 85956 4172 2808144 0 0 58496 49708 1570 2087 3 33 64 0
2 1 0 84792 4172 2809344 0 0 58112 51992 1585 2058 3 27 70 0
2 1 0 86176 4172 2807836 0 0 58240 52836 1599 2029 1 29 70 0
2 1 0 86164 4172 2803968 0 0 62080 53500 1664 2085 2 40 58 0
2 1 0 85152 4172 2804940 0 0 57600 51320 1572 1837 3 37 60 0
1 1 0 85840 4172 2804592 0 0 58752 52500 1603 1917 2 48 50 0
1 1 0 85760 4128 2804836 0 0 58368 51100 1577 1853 3 35 62 0
0 2 0 85456 4128 2804532 0 0 3712 45740 675 289 0 6 94 0
1 1 0 84796 4128 2805896 0 0 30592 47528 1113 1047 0 27 73 0
1 1 0 85708 4132 2808376 0 0 52356 43612 1422 1755 0 26 74 0
1 1 0 85852 4132 2808212 0 0 58368 46252 1528 1880 2 35 63 0
1 1 0 86120 4136 2807972 0 0 50052 46212 1411 1740 3 30 67 0
1 1 0 82728 4136 2808276 0 0 57472 50500 1560 1866 4 29 67 0
3 1 0 83316 4140 2807428 0 0 56324 44780 1490 1875 2 39 59 0
2 1 0 82332 4140 2807960 0 0 59648 48724 1578 1905 0 34 66 0


time bash -c "dd if=~mmokrejs/video/kamcatka2004.dv of=/mnt/ext2/kamcatka2004.dv bs=1024 count=4096000; sync"
4096000+0 records in
4096000+0 records out

real 1m29.880s
user 0m1.600s
sys 0m12.220s

or in another attempt:

real 1m33.282s
user 0m1.430s
sys 0m12.180s


on 2.6.13-rc6-git9 (read from local xfs disk to another disk with ext2 fs):

procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
r b swpd free buff cache si so bi bo in cs us sy id wa
1 10 0 85452 7628 2754552 0 0 57984 52188 876 1647 2 36 0 62
0 10 0 85452 7380 2756932 0 0 59520 46260 862 1620 1 41 0 58
1 9 0 85080 7424 2756564 0 0 53636 46188 838 1515 1 34 0 65
1 9 0 85328 7476 2755472 0 0 59400 51077 897 1737 1 39 0 60
1 10 0 85204 7528 2754940 0 0 53508 51456 844 1513 2 33 0 65
0 10 0 85452 7580 2753804 0 0 58368 50956 879 1682 2 33 0 65
0 10 0 85204 7636 2753180 0 0 56196 51512 869 1597 2 33 0 65
0 10 0 85204 7688 2752176 0 0 61696 45856 906 1728 1 36 0 63
1 9 196 85080 7740 2753368 0 0 54276 51456 848 1572 1 40 0 59
0 10 196 85328 7784 2752968 0 0 59008 30752 888 1658 1 40 0 59
0 10 196 85328 7820 2753376 0 0 58368 23520 883 1659 1 33 0 66
1 10 196 85452 7856 2753016 0 0 58756 48164 922 1646 1 37 0 62
0 10 196 85328 7252 2754324 0 0 57600 49172 878 1609 1 37 0 62
1 9 196 85204 7296 2754564 0 0 56708 43176 853 1598 2 36 0 62
1 10 196 85204 7324 2754496 0 0 59392 47600 882 1674 0 41 0 59
0 10 196 85328 7288 2754280 0 0 54660 46200 831 1531 1 34 0 65
1 9 196 85452 7224 2753856 0 0 59008 45808 876 1657 2 37 0 61
0 10 196 85080 7164 2753880 0 0 57476 46144 865 1619 2 34 0 64
0 10 196 85204 7116 2753480 0 0 56968 46209 860 1658 1 32 0 67

# time bash -c "dd if=~mmokrejs/video/kamcatka2004.dv of=/mnt/ext2/kamcatka2004.dv bs=1024 count=4096000; sync"
4096000+0 records in
4096000+0 records out

real 1m31.067s
user 0m0.760s
sys 0m15.809s

and another trial:

real 1m31.266s
user 0m0.796s
sys 0m16.361s


>>The throughput is clearly lower on 2.6 kernel and definitely the
>>CPU is in my eyes unnecessarily blocked... Why is the CPU in the
>>wait state instead of idle (this is teh problem on 2.6 series
>>but CPU is free on 2.4 series)? That's the main problem I think at the
>>moment.
>
>
> There is no wait state accounted for in 2.4 so you won't see it.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/