| Tweetovi |
|
Glauber Costa
@glcst
|
31. sij |
|
So it's left as an exercise for the reader.
|
||
|
|
||
|
Glauber Costa
@glcst
|
31. sij |
|
In 5.7 ?
|
||
|
|
||
|
Glauber Costa
@glcst
|
31. sij |
|
take a look at the supported opcodes for 5.6. You can't do any syscall, that is true, but there are lots of non-io stuff.
|
||
|
|
||
|
Glauber Costa
@glcst
|
31. sij |
|
"It is so good it might morph into a general asynchronous system call interface." might, lol
|
||
|
|
||
|
Glauber Costa
@glcst
|
31. sij |
|
I've been dealing with SQL and CQL for so long, that I know instinctively read the names of the io_uring data structures as "structured completion entry" and "cassandra completion entry". #JustSaying
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
Gosh! When was that ?
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
Say goodbye to Linux. The kernel is now called bpf_uring
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
Getting started with io_uring for sockets for seastar. Good that I am doing this per-merge of the storage bits, because some design decisions will have to change to accommodate both more elegantly. Soon seastar will need only one system call #io_uring
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
why ? what is wrong with it ?
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
impeach build systems
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
impeach cmake!
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
cmake stole my soul. You can't steal anything further of value.
|
||
|
|
||
|
Glauber Costa
@glcst
|
30. sij |
|
I can't really control this, I am not informed enough to make these kinds of decisions in the project. I have really only ever needed to deal with cmake twice so far, and what I can say is that I would be willing to migrate to literally anything else including hell itself.
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
I lost (am losing) so much time with cmake these past days, it's not even funny
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
yes, although it is rare for Scylla. For seastar, the framework, though, people use it for god knows what so I like to test the extremes
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
already did, waiting for his reply. BTW, I can send patches to liburing to io-uring@vger too, right ? Wanted to CC @AviKivity on that one, it was his idea)
Guess I fired first and asked later, but that's standard me
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
that's the workload I expect the biggest bang for the buck, though (because of the competing CPU workload). Without the competing CPU hog I got better, but by about 5%. Both aio and uring can easily max out my nvme, though, so now I am looking for a better one
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
Wondering how are the early results for the io_uring backend for seastar? 50% faster in the first benchmark (workload is small 512-byte reads with iodepth of one, competing for dispatch time against a CPU-bound constant workload)
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
Indeed, which is why it is so important for their infrastructure to work well. If it doesn't, all we're left with is political fights on twitter. Discord is necessary for world peace.
|
||
|
|
||
|
Glauber Costa
@glcst
|
29. sij |
|
Discord is migrating from Cassandra to Scylla. It caused them high latencies because they had to read more from their current Cassandra cluster to siphon the data. Easily solved by slowing down. But wait until you see how low those latencies will be once the migration finishes! twitter.com/AviKivity/stat…
|
||
|
|
||