Blog‎ > ‎

Double bind, I/O latency, /tmp tmpfs, Social loafing, WhatsApp, Social Credit

posted Jan 15, 2017, 2:16 AM by Sami Lehtinen   [ updated Jan 15, 2017, 2:27 AM ]
  • Double bind - Aah, excellent topic. World is just so full of double binds. Every project and almost every task contains major contradictions how things should be done. This topic is so highly depressing, I'm not even starting about it. This is a subject for several books. - Also the new Finnish Spy law will be naturally full of this stuff. You must do it, but you must not do it, etc. Because utilizing this right would violate rights of others etc. - Actually this is quite closely related to the 'everything is a trade-off' thinking. There's no absolute right / correct solution to many problems.
  • Wrote a small test application which measures time of 4k file creation and stores it. This is perfect method for monitoring underlying storage systems performance. As well as doing random 4k reads (slowly) in the space of whole storage system. This has been very beneficial. It's easy to see that small number of storage writes seem to be extremely low, compared to average. But as said, this shouldn't be anything new to storage / performance / cloud people. Generally it works, but at times, it's just extremely slow. Like 1000000 times slower than it's supposed to be. This can be caused of course by multiple reasons. I personally think that the primary source has some kind of problem, and it causes timeout and then operation is tried on some other source, which might as well timeout. Leading to cascading of timeouts, which leads to extremely low compared to normal operation.
  • Run base line tests on HDD and SSD storages as well as on multiple cloud storages. It seems that with Cloud Storage it's highly likely you'll be getting consumer spinning rust kind of performance. Instead of screaming fast SSD / Enterprise server performance. - Of course everyone's expecting that cheap systems would be extremely fast enterprise systems, right? - At least I'm not expecting that. But still, latencies need to be reasonable. As well as single thread blocking programs severely suffer if I/O latencies grow.
  • Fsyncing on /tmp still writes to underlying storage. I actually didn't know that. I thought it would be RAM stored, and only if there's memory pressure it would get flushed to underlying storage. Yet, after I examined this stuff bit more on the test computer. I did find the reason. I assumed /tmp/ would use tmpfs, but it doesn't on Debian & Ubuntu. So that's the reason why it still wrote to disk. Changed fstab so that /tmp/ is now stored on tmpfs. /etc/fstab "tmpfs /tmp tmpfs nodev,nosuid,noexec,nodiratime,size=1G,mode=1700   0 0" Unsurprisingly the tmpfs based /tmp is also a bit faster.
  • Endless discussions about storage latency. Now they claim that the test software is reason for the latency. That's BS. I always like to check affecting factors and I can tell that on ramdisk the 4k file creation latency is around 0.6 ms on my desktop. So anything above 1 ms is guaranteed to be caused by the storage subsystem.
  • Wanted to watch the ExoMars and TGO & Schiaparelli landing via live stream. But the stream kept hanging all the time. Sigh. So typical, we live stream this and then the live stream won't work.
  • Social loafing - Just so normal. Everybody thinks that somebody will take care of it. Haha, so classic and traditional.
  • EFF's article about WhatsApp security. Interestingly they had exactly the same points I've mentioned earlier. Having top notch all the time changed encryption keys is totally meaningless, if there are multiple circumvention methods left in place.
  • In some earlier posts I've said that trust and reputation is always context related. But in China they seem to think that the trust can be widened. Generally if you're 'being trusted and keep your word' it usually means that you could be trusted on other aspects too? Or does it? It depends. Arguments for both directions can be made and of course examples found. But this is what China is planning a National 'trust / social credit' score. In a way that's good, because now breaching trust on some level, will also cause issues on others. Which will make you less likely to want breach trust in cases where you think it just won't matter even if I don't keep my word or do the right thing. In many important cases the generic integrity of your word is counted more important than what you actually did. If they ask if you have done something you promised to do. Well, of course it's bad if you haven't. But if you then lie about it, it's even worse. Then it's immediately clear that you can't be trusted on even really small things. So how you could be trusted on more important things? In Finland trust and reputation is highly valued. In other countries on cultures it seems that it doesnt't matter, if you got really bad reputation. Yet continuous lack of trust makes many things really hard to organize. Isn't it great if you can be sure that the counter party got probity a complete and confirmed integrity. It's much better than just a plain credit score. In Finland people don't have credit score. Everyone's "trusted by default", you'll only lose your credit score if you screw up. It's like having or not having a criminal record. Even if we know that practically everyone has broken the law. Of course if that kind of system is being abused, it would be pretty horrible. It's like the double construction cases. If the 'military intelligence' spots something which isn't right, but isn't about military, should they rat on it? Isn't that still abuse of power, or is it just for generic greater good? Don't know, there are hard things to make right conclusions about. Finland it's right now crafting laws about these issues. Do you need privacy, if you don't do anything wrong? If the lack of privacy is handled correctly and not abused, does it matter and so on. If we accept that 'military intelligence' or 'state security agency' or whatever, got basically full access to everything. Is that a problem? It shouldn't be or it could be? Who knows.