Brian Candler
2/10/2007 1:56:00 PM
On Sat, Feb 10, 2007 at 07:42:45AM +0900, Alex Young wrote:
> I thought I understood what user and system times meant. Then I saw
> this from Benchmark, while comparing response times of a WEBrick server
> and a Mongrel server:
>
> user system total real
> webrick 29.580000 4.920000 34.500000 ( 87.191739)
> mongrel 27.640000 4.540000 32.180000 ( 58.508197)
>
> Clearly the mongrel server wins, but why does it only show in the
> wallclock measurement?
With webrick, the process is waiting for some external event.
> Where did the rest of the time go? This is
> repeatable, so I don't think it's interference from any other processes
> on the machine.
Something which might be hurting webrick is if it hasn't turned off TCP
Nagle (socket option TCP_NODELAY). If the process sends less than 1500
bytes, the kernel will wait for 0.1 seconds to see if there's more to come
before actually sending a packet.
However, in the real world this is unlikely to be a problem.
> To be clear here, what I'm actually measuring is the time taken for
> 10000 queries to be serviced from a separate process, and the webrick
> and mongrel servers are also each in their own process. Everything's on
> localhost.
You'd need to be more specific than that. Are you opening a fresh TCP
connection for each query? Or are you sending multiple queries down the same
connection, using HTTP/1.1?
Are you sending 10000 queries one after the other, or do you have (say) 100
query threads, each sending 100 queries?
30 seconds of wallclock time averages only 3ms per query, but the Nagle
explanation might work if you are running, say, 32 queries concurrently.
Other possibilities might be writing to log files, if the process does an
open-write-close or write-flush each time. Running the process under
'strace' might give you a better idea.
HTH,
Brian.