From: Rodrigo C. <rn...@gm...> - 2007-03-05 03:00:00
|
Yes, all that is correct, and that's why I told you we should have choices. The API should provide choices and flexibility. Decent programmers will use the right tool, not the wrong tool. Documentation should also help whenever possible. BTW Mark, does my API solve your random-access problem?... Jimmy Zhang wrote: > Totally! DRAM access time is 100~200 cycles, packing everything together > will make cache miss far less often and boost both parsing and > navigation perfomrnace! > ----- Original Message ----- From: "Mark Swanson" > <ma...@Sc...> > To: "Rodrigo Cunha" <rn...@gm...> > Cc: "Jimmy Zhang" <cra...@co...>; > <vtd...@li...> > Sent: Sunday, March 04, 2007 10:35 AM > Subject: Re: [Vtd-xml-users] Random Access Proposal (take 2) > > >>> Memory usage is not allways critical, sometimes performance can be >>> gained from memory, sometimes that's a false argument. In this case an >> >> Just my 2cents: significant performance increases can be gained by >> structuring your algorithms to keep your memory accesses within a >> column of memory as much as possible. DRAM cycle time penalties are >> huge - unless you can keep the data in the CPU's cache. >> This is one of those tricks that column databases like K and Vertica >> use to stomp traditional databases with for n-dimensional queries. I >> actually wrote an n-dimensional column-based database for a company a >> number of years ago. >> >> Cheers. >> >> -- >> http://www.ScheduleWorld.com/ >> Free Google Calendar synchronization with Outlook, Evolution, >> cell phones, BlackBerry, PalmOS, Exchange, Mozilla, Thunderbird, >> Pocket PC/Windows Mobile. Also sync tasks, notes and contacts! >> WebDAV, vfreebusy, RSS, LDAP, iCalendar, iTIP, iMIP support. >> > > > |