On 17/09/2012, at 06:35, Ben Noordhuis wrote:

> (...)
> That small gain is offset by the need to serialize access to your data
> structures - which is often a lot more expensive than the small gain
> you get by using threads.
> 
> As for transferring objects, you don't need threads for that, just
> shared memory.

If the processes are sharing memory, then they *too* "need to serialize access 
to data structures"...

> We'll probably implement that someday but don't expect too much from
> it. Shared memory avoids some syscalls but the cost of moving the
> object from one V8 heap to another remains.

That's the problem for transferable objects: there's no way to grab an object 
reference from isolate A to use it on isolate B.

> By the way, if you want to hasten that day, post (non-contrived)
> benchmarks that conclusively show that IPC is a bottleneck. :-)

If the processes can communicate via shared memory -which is always a given for 
threads- then IPC is fast.

But if they can not then you've got to copy the data and speed becomes a 
function of ( data.length ) which might be *irremediably* slow.

Big data.length copies also flush other data from the caches, which results in 
extra slowdowns.

And as the memory bus is a shared resource, under high loads these (many) 
unnecessary big.data.length copies will (pretty soon) have a global impact on 
the performance of *all* the rest of system (รก la `cat /dev/zero > /dev/null` 
memory bus bandwidth exhaustion).
-- 
Jorge.

-- 
Job Board: http://jobs.nodejs.org/
Posting guidelines: 
https://github.com/joyent/node/wiki/Mailing-List-Posting-Guidelines
You received this message because you are subscribed to the Google
Groups "nodejs" group.
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/nodejs?hl=en?hl=en

Reply via email to