Arnaldo, Leandro, - | >> For the BAD and for last GOOD using HZ=250 I got: | >> | >> "warning: #warning Coarse CONFIG_HZ resolution -- higher value | >> recommended for TFRC." | >> | >> Leandro. | >> | > | > BAD: Using CONFIG_HZ=1000 "Protocol Not attached" | > GOOD: Using CONFIG_HZ=1000 "OK" | > | | Tests using High Resolution Timers using BAD: | | 100hz OK | 250hz OK | 1000hz OK | This confirms that the cause of the problem was in not enabling High-Resolution timers. There will then be the warning message in the system logs saying that the timer resolution is too low. I tried to find Leandros .config file again, it seems that high-resolution timers were not enabled; the only other possibility is when passing clocksource=jiffies (or something else coarse-grained). The HZ variable is not really relevant here; you are still seeing this messages since your version of the test tree is older than 2 weeks: http://marc.info/?l=dccp&m=122451741426680&w=2 The fact that the CCID-3 module is not loaded when the timer resolution is too coarse is not a bug, it is there to avoid problems that result when using coarse-grained timers on high-speed interfaces: most PCs laptops now come with Gigabit Ethernet per default, local loopback also operates at high speeds; when using such interfaces I noticed several problems as described in the commit message, http://eden-feed.erg.abdn.ac.uk/cgi-bin/gitweb.cgi?p=dccp_exp.git;a=commitdiff;h=42c46151aedc99689341ac73131cd56c2aec55c4 Maybe it is a good idea to rethink the whole high-resolution timer issue. I can see only two ways: (1) Consistently use coarse-grained timers everywhere (jiffy resolution) -------------------------------------------------------------------- Ian McDonald suggested this, and I can see the advantages - we could unify RTT measurement for all the CCIDs, use a simpler implementation. (2) Consistentlly use high-resolution timers everywhere --------------------------------------------------- This would require to use the above protection in general. In a discussion about the build warning, Dave Miller suggested converting to high-resolution timers and he is right about it - if we decide to stick with high-resolution timers - since mixing high-resolution code (currently the computation of X_recv and RTT) with low-resolution code (dccp_xmit_timer) is neither here nor there and gives bad performance. Arnaldo, Leandro and others - it would be good to have input on this. I have looked into the conversion to high-resolution timers, but I have doubt that it will improve the performance of CCID-3. >From earlier discussion I know, Arnaldo, that you were in favour of keeping the resolution high. But if we want to do that then I think we need to convert the rest to high-resolution timers also. Or decide to use (1). Gerrit -- To unsubscribe from this list: send the line "unsubscribe dccp" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
