RE: Why use threads ( was: Alan Cox quote?)
Davide Libenzi (davidel@xmailserver.org)
Wed, 20 Jun 2001 19:22:45 -0700 (PDT)
On 21-Jun-2001 David Schwartz wrote:
>
>> > Who said anything about 'select'? If you want to learn
>> > how to write
>> > efficient multi-threaded servers, take a course or read a book.
>> > Heck, you
>> > can even ask me questions on marginally appropriate lists or
>> > even by private
>> > email. But don't put words in my mouth.
>
>> I was just thinking about having a course on how to write mt applications,
>> are You currently keeping such courses ?
>
> I've been preparing one for several years, but due to the constantly
> changing state of the art and the other limitations on my time, it just
> keeps getting further behind.
>
>> Is still this Your address :
>>
>> David Schwartz
>> 16000 NW Modesty Dr
>
> That was never my address, though I lived at 16000 NW 1st street a long
> time ago.
>
>> How do you handle an average of 1600 sessions over a single
>> process without
>> using select()/poll(), I'm just curious ?
>
> Well, with 1,600 connections, things are pretty easy. This is so far
below
> the limit of modern machines that efficiency only matters if your server is
> just one of many things the machine does. I would just use two threads in
> poll loops, each working on half the descriptors. Some would have these
> threads actually do the I/O, others would have it queue I/O jobs to another
> pool of I/O threads that do the actual read/write operations.
>
> My (WebMaster's) library does even better than this, converting the
'poll'
> threads into 'do the I/O' threads dynamically. That way if the 'poll' only
> hits on one file descriptor, you don't have to do a context switch to
> service the I/O, but you also can get back to 'poll' pretty quickly even if
> the I/O manages to block when it's not supposed to.
>
> But 1,600 is easy, so there's no reason to sweat about it.
>
> Things get more difficult at 16,000 connections. At this level, I
recommend
> a tiered approach. Separate the file descriptors into the 80% that are 20%
> of the activity and find the 10% that are 90% of the activity. Have separate
> threads poll on each of these groups. The advantage of this is that the more
> expensive poll calls (the ones on the greatest number of file descriptors)
> are called very rarely (because those file descriptors aren't very active.
> Tracking code can move file descriptors dynamically from group to group.
>
> No matter what anyone tells you, 'poll' scales *better* than O(n) (in
other
> words, the more connections you have, the less CPU time you need per
> connection to discover which sockets need work), and since your I/O can't
> possibly scale better than O(n), poll is as scalable as it needs to be. If
> you double the number of sockets, you double the cost of 'poll' but you also
> double the amount of information you get per poll call (actually, you more
> than double it, but that's a long story).
>
> The problem with 'poll' is efficiency at *low* load. Since I write
mostly
> servers designed to operate at high load, I don't worry too much about
> efficiency at low load. The hard case for 'poll' is large numbers of file
> descriptors at very low load (so you're unlikely to find more than one
> 'active' fd at a time). Fortunately these cases don't need much efficiency.
> The operating systems max out at around 65,536 descriptors anyway, and
> keeping these inactive enough to allow such low discovery rates means a
> server with most of its CPU to spare.
>
> Not that I have anything against the more efficient I/O discovery
> techniques under discussion and development. There's nothing wrong with a
> more efficient approach, especially one that's more efficient at every
> combination of loads and socket counts. But as far as ultimate scalability
> goes, socket discovery is not the limiting factor -- far from it.
Just to summarize :
1) You said to handle 16000 sessions with 10 threads
2) I said: "Humm, you've to select() about 1600 fds ..."
3) You said : "Who said anything about 'select'?"
The stuff above looks like select() to me.
I assume You know how select is implemented inside the kernel
( file->f_op->poll() ), so I don't understand Your phrase :
<quite rude>
Who said anything about 'select'? If you want to learn how to write
efficient multi-threaded servers, take a course or read a book.
</quite rude>
About the scale factor of select()/poll() my agreement is only partial.
Have You ecer observed that Your processes tend to become a bit CPU bound when
stocking a lot of fds inside a poll ?
- Davide
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/