<br /><br /> Lamar Owen wrote:<br /><blockquote cite="mid:200204141515.39014.lamar.owen@wgcr.org"
type="cite"><blockquotetype="cite"><pre wrap="">The low-latency and preemptible patches are not meant for
performance<br/>gains, but for responsiveness, and are not designed to be used in servers,<br />only in
workstations/desktops.<br/></pre></blockquote><pre wrap=""><br />ISTM that improving interactive performance would also
improvemultiuser <br />performance in a server, as low latency and kernel preemption can increase <br />multiuser
serverresponsiveness.</pre></blockquote> responsiveness != performance IT works OK for a low number of concurrent
users/processesto increase percieved performance, but to get real gains on large systems with large numebrs of users
andprocesses you actually decrease the responsiveness of individual tasks (IE make the system a little less likely to
contextswitch or pre-empt) and schedual in batches or clusters rather than one-at-a-time. For a desktop/workstation
thiswould be insane, and drive a user to kill someone, but for systems that handle several hundred users (interactive
ornot) this improves overall perfomance.<br /><br /> 2.4.18 has a lot of work done to the VM, but most importantly has
workdone to the queue elevator code, thats probably whats doing most of the work (throttling big writers) of seeing
betteroverall system performance.<br />