isn´t this weird:
# time foo real 43m39.841s user 15m31.109s sys 0m44.136s
Almost 30 minutes have disappeared, but it actually took about that long, so what happened?
This is counting the CPU time that a process used. If something is not in ‘CPU’ but waiting on input etc it might not get counted in user or sys. There is also the fact that the builtin bash time command you used calculates things differently from the /usr/bin/time command.
From the /usr/bin/time man page
Note: some shells (e.g., bash(1)) have a built-in time command that
provides less functionality than the command described here. To access
the real command, you may need to specify its pathname (something like
From the bash man page
When the shell is in posix mode, time may be followed by a newline. In
this case, the shell displays the total user and system time consumed
by the shell and its children. The TIMEFORMAT variable may be used to
specify the format of the time information.
The built in time is actually meant to be used with measuring pipeline information but can be used by itself
Stephen J Smoogen.
I may misunderstand your question, but
“time” is provided by the bash shell. It may be provided by a command if you are using a different shell. When the command following the
“time” keyword completes, bash will print the amount of elapsed time
(the amount of time that passed between the command’s start and its exit), the amount of time the command was using the CPU and not in a sleep state, and the amount of time the kernel was using the CPU to service requests from the command.
So your “foo” application was in a sleep state for around 30 minutes of the 44 minutes that passed between when you started it and when it finished.
Gordon Messmer wrote:
Hm. Foo is a program that imports data into a database from two CVS files, using a connection for each file and forking to import both files at once.
So this would mean that the database (running on a different server) takes almost two times as much as foo — which I would consider kinda excruciatingly long because it´s merely inserting rows into two different tables after they were prepared by foo and then processes some queries to convert the data.
The queries after importing may take like 3 or 5 minutes. About 4.5 million rows are being imported.
Would you consider about 20 minutes for importing as long?
That depends on a lot of things.. from drive speed to drive layout to database to network congestion to… without that information the question is not answerable.
There are far too many variables you’ve not mentioned to determine if that’s good or bad (or very bad). Is the connection a local connection
(ie the import is done on the DB server) or a network connection?
What size are the CSV (CVS is a typo, correct?) files? 4.5M rows tells us nothing about how much data each row has. It could be 4.5M rows of one INT field or 4.5M rows of a hundred fields.
I’m a bit confused by the last two sentences. Based on how I read this:
1. Foo is prepping (creating?) the tables
2. Processes queries to convert the data (to CSV?)
3. Runs more queries on those tables.
Or it could be:
1. Foo preps the tables
2. Foo imports the CSV files
3. Foo does post-processing of the tables.
It’s not really clear the actual process, but I’ll go on the assumption that Foo is creating the tables with the correct fields, data types, keys and hopefully indices. Then dumps the CSV files into the tables. Then does post-processing. (I’ve written similar scripts, so this is the most logical process to me.)
If we assume network bandwidth is fine, that still leaves far too many server variables to know if 20m is about right or not. Amount of data to import, TYPE of data, database AND server configuration, CPU, RAM, etc and DB config for tunable paramters like buffer pool, read/write I/O
IIRC, you posted some questions about tuning a DB server a while back, would this be data going into that server, perhaps?
I’d like to offer a helpful suggestion when asking for list help. It’s better to provide TOO MUCH information, than too little. There’s a big difference between ‘my printer won’t print’ and ‘my printer won’t print because it’s not feeding paper properly’.
Mark Haney Network Engineer at NeoNova
919-460-3330 option 1
so you’re missing about 25 minutes, and maybe 5 minutes is spent post processing, so thats 20 minutes spent in the data insertion?
inserting one row at a time? or in batches? remeber a database server is going to do commits after each transaction, which forces the data to be flushed to disk. 4.5 million seperate row transactions, yeah, I could see that taking some time, plus add that many network round trips, etcetc. if the db server just has a single SATA disk, you’re doing 9 million committed writes combined to the two tables?
20 minutes for 9 million inserts, thats 7500 per second.
john r pierce, recycling bits in santa cruz
Mark Haney wrote:
Foo is running on a different machine than the database server.
One CSV is 70745427, the other one is 536302424 bytes (68M and 512M). That´s 18 and 23 fields or so to insert for each row.
… deletes the part of the rows that was imported the last time. The rows from last time are being imported again, plus new rows.
Only importing new rows would require checking every row that is being imported to figure out if it´s already there, which may not be so much faster as to be worthwhile, and since I usually don´t need to wait on the import to finish, it doesn´t really matter.
The servers are connected by 4x1GB, using LACP.
Of course — what gives me to think is that it takes relatively long for the database to insert the rows while foo converting them is relatively fast.
Foo is written in perl. I like to think that letting the database do as much of the work as possible is generally a better idea than doing things that the database could do in perl because the database is likely to be faster — without overdoing either because for practical reasons, things need to be kept sufficiently simple, and unnecessary optimization is, well, unnecessary.
Now I wonder if my general assumption is false, though foo isn´t a good example to verify the assumption because it can´t really do anything else but import the rows, which takes as long as it takes, and the post processing is surprisingly fast (and brings the time that queries take which are working with the data once it has been imported down from many hours to a few minutes or to seconds and less because I optimized things).
So I don´t know … I guess 45 minutes to import 600MB of data is reasonably fast, considering that 2.25 million rows times 40 fields yield 90 million fields, so that´s about 3333 fields/sec.
John R Pierce wrote:
Yes, with the 15 minutes actually spent on foo spent on converting the fields and sending them to the server, which I think is pretty good.
They are inserted one row at a time, during one transaction for each of the CSV files. I´d have to figure out how to insert them in batches, that might yet be faster. I could easily stack up 1000 rows or so and then insert them all at once, if that´s possible.