HI !
I'm very new to L4Linux so my question might be obvious - anyway did not find anything directly related in the archive. For a start I did a quick performance comparison using lmbench-3.0.
This comparison is based on the L4Linux-2.6 running a debina ramdisk lmbench-3.0 vs Linux-2.6.14 running from harddisk (though data files were in tempfs in both cases - so the HD should not really have much of an impact).
L4Linux Linux 2.6.14
com lat.: (us) lat_pipe 14.1-14.5 2.0 lat_fifo 13.9-14.6 2.2 lat_unix 19.4-19.5 3.2 lat_sem 3.5 0.6
lat_proc: (us) fork 805-1249 80-82 exec 3124-3784 212-230 shell 9992-9997 2480-2592
tlb: 7,8,32 always 32
bw com: (MB/s) bw_pipe 921-958 1338-1352 bw_unix 429-753 1122-1138
lat mem: oopses under L4Linux - no results
bw_mem: (MB/s) bcopy 3856-4066 5056-5080 rdwr 11937-13242 16052
syscall: (us) null 2.4-2.8 0.13 read 2.6-3.3 0.26-0.31 stat 4.4-4.9 0.89 open 6.5-8.1 1.16-1.39
The thing that supprises me is that the performance variation is much higher than on regular linux - in both cases the system was idle except for lmbench running, so I'm excluding any syslog parties or cron jobs fireing at inconvenient times (i.e bandwidth of UNIX sockets has a variance of almost 50% ?). Generally the performance difference is much larger than I would have expected, especially also in areas where I don't quite understand where the u-kenrel can impact the Linux system (i.e. bcopy, memory read-write)
Are these values in the usual range or did I screw up fundamentally ?
thx ! Prof. Nicholas Mc Guire University of Lanzhou Distributed & Embedded System Lab http://dslab.lzu.edu.cn School of Information Science and Engeneering mcguire@lzu.edu.cn Tianshui South Road 222. Lanzhou 730000 .P.R.China Tel:+86-931-8912025 Fax:+86-931-8912022 Lichtensteinstr. 31, Mistelbach A-2130 Austria Tel:+43-2572-201082 Fax:+43-2572-201084
Dear Prof. Mc Guire,
On Wednesday 07 December 2005 21:39, Der Herr Hofrat wrote:
I'm very new to L4Linux so my question might be obvious - anyway did not find anything directly related in the archive. For a start I did a quick performance comparison using lmbench-3.0.
[...]
The thing that supprises me is that the performance variation is much higher than on regular linux - in both cases the system was idle except for lmbench running, so I'm excluding any syslog parties or cron jobs fireing at inconvenient times (i.e bandwidth of UNIX sockets has a variance of almost 50% ?). Generally the performance difference is much larger than I would have expected, especially also in areas where I don't quite understand where the u-kenrel can impact the Linux system (i.e. bcopy, memory read-write)
Are these values in the usual range or did I screw up fundamentally ?
L4Linux 2.6.14 is still a bit experimental. At least it still did not get _that_ level of optimization that L4Linux 2.0 from the original SOSP'97 paper got. I assume that you used Fiasco as the underlaying L4 microkernel. The development of that kernel had mainly good preemptablility in mind and it is therefore not that performance tweaked as other L4 kernels.
It would be nice to get some more words about your hardware and software configuration, e.g. which CPU and how many RAM, the configuration of Fiasco and L4Linux. There are several points to optimize.
Frank
l4-hackers@os.inf.tu-dresden.de