diff options
Diffstat (limited to 'open_issues/multiprocessing.mdwn')
| -rw-r--r-- | open_issues/multiprocessing.mdwn | 37 |
1 files changed, 34 insertions, 3 deletions
diff --git a/open_issues/multiprocessing.mdwn b/open_issues/multiprocessing.mdwn index 224c0826..562ccd83 100644 --- a/open_issues/multiprocessing.mdwn +++ b/open_issues/multiprocessing.mdwn @@ -1,4 +1,4 @@ -[[!meta copyright="Copyright © 2010 Free Software Foundation, Inc."]] +[[!meta copyright="Copyright © 2010, 2011 Free Software Foundation, Inc."]] [[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable id="license" text="Permission is granted to copy, distribute and/or modify this @@ -8,7 +8,7 @@ Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled [[GNU Free Documentation License|/fdl]]."]]"""]] -[[!tag open_issue_hurd]] +[[!tag open_issue_documentation open_issue_hurd]] We would expect that fine-grained, compartmentalized systems, that is, microkernel-based multi-server systems in particular, would be ideal candidates @@ -16,7 +16,7 @@ for applying multiprocessing. That is, however, only true from a first and inexperienced point of view: there are many difficulties. -IRC, #hurd, August / September 2010 +IRC, freenode, #hurd, August / September 2010 <marcusb> silver_hook: because multi-server systems depend on inter-process communication, and inter-process communication is many times more @@ -31,6 +31,37 @@ IRC, #hurd, August / September 2010 serious research challenges +IRC, freenode, #hurd, 2011-07-26 + + < braunr> 12:03 < CTKArcher> and does the hurd take more advantages in a + multicore architecture than linux ? + < braunr> CTKArcher: short answer: no + < CTKArcher> it's easier to imagine one server pro core than the linux + kernel divided to be executed on multiple cores + < braunr> CTKArcher: this approach is less efficient + < braunr> CTKArcher: threads carry state, both explicit and implicit (like + cache data) + < braunr> CTKArcher: switching to another core means resetting and + refetching this state + < braunr> it's expensive and there is no gain obtained by doing this + < braunr> thread migration (having a thread from a client also run in + servers when making synchronous RPC, even handling its own page faults) + was implemented in mach4 and is imo a very good thing we should have + < braunr> CTKArcher: and concerning linux, it's actually very scalable + < braunr> it's already like if all client threads run in servers (the + kernel is the servers there) + < braunr> rcu is used a lot + < braunr> thread migration already takes into account smt, cores, and numa + < braunr> it's hard to do something better + < braunr> (here, thread migration means being dispatched on another cpu) + < braunr> some systems like dragonflybsd go as far as to pin threads on one + processor for their entire lifetime + < braunr> in order to have rcu-like locking almost everywhere + < braunr> (you could argue it's less efficient since in the worst case + everything runs on the same cpu, but it's very unlikely, and in practice + most patterns are well balanced) + + debian-hurd list On Thu, Jan 02, 2003 at 05:40:00PM -0800, Thomas Bushnell, BSG wrote: |