From: Christoph Lameter <clameter@sgi.com>
To: Robin Holt <holt@sgi.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>,
"Luck, Tony" <tony.luck@intel.com>, Adam Litke <agl@us.ibm.com>,
linux-ia64@vger.kernel.org, torvalds@osdl.org,
linux-mm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: Increase page fault rate by prezeroing V1 [3/3]: Altix SN2 BTE Zeroing
Date: Wed, 22 Dec 2004 11:56:47 -0800 (PST) [thread overview]
Message-ID: <Pine.LNX.4.58.0412221154340.9186@schroedinger.engr.sgi.com> (raw)
In-Reply-To: <20041222124647.GB1813@lnx-holt.americas.sgi.com>
I have done some additional tests with a 128 cpu SMP machine and it shows
that the bte slows things down during memory benchmarking by about 10-20%
although its causing less load when the system is not under high stress.
So its not always win and I may drop bte support in a future version. Can
we talk off list about this since this is mostly an SGI thing?
On Wed, 22 Dec 2004, Robin Holt wrote:
> We still need to talk. This is a much smaller patch, which I like. The
> problem I see in my 30 second review is you are doing things per-cpu
> when they really need to be done per-node. It is very likely that
> there will be M-Bricks in the system (cranberry2 has one if you want
> to test your code out there or you can take any altix and disable the
> cpus on a C-Brick). With M-Bricks, you will essentially limit
> yourself to one zero operation per controlling node instead of one
> per node.
>
> I think the easy answer is to not have the structure allocated
> within bte_copy(), but rather within bte_start_zero and passed
> in as the notification address.
>
> Give me a call sometime today (Wed. I am in the office from about
> 10:00 CDT until around 4:00 CDT) Maybe we can get this straightened
> out quickly. If you are not calling from the office, email me with
> other arrangements.
>
> Thanks,
> Robin
>
> On Tue, Dec 21, 2004 at 11:57:57AM -0800, Christoph Lameter wrote:
> > o Use the Block Transfer Engine in the Altix SN2 SHub for background zeroing
> >
> > Index: linux-2.6.9/arch/ia64/sn/kernel/bte.c
> > ===================================================================
> > --- linux-2.6.9.orig/arch/ia64/sn/kernel/bte.c 2004-12-17 14:40:10.000000000 -0800
> > +++ linux-2.6.9/arch/ia64/sn/kernel/bte.c 2004-12-21 11:03:49.000000000 -0800
> > @@ -4,6 +4,8 @@
> > * for more details.
> > *
> > * Copyright (c) 2000-2003 Silicon Graphics, Inc. All Rights Reserved.
> > + *
> > + * Support for zeroing pages, Christoph Lameter, SGI, December 2004.
> > */
> >
> > #include <linux/config.h>
> > @@ -20,6 +22,8 @@
> > #include <linux/bootmem.h>
> > #include <linux/string.h>
> > #include <linux/sched.h>
> > +#include <linux/mm.h>
> > +#include <linux/scrub.h>
> >
> > #include <asm/sn/bte.h>
> >
> > @@ -30,7 +34,11 @@
> > /* two interfaces on two btes */
> > #define MAX_INTERFACES_TO_TRY 4
> >
> > -static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
> > +DEFINE_PER_CPU(u64 *, bte_zero_notify);
> > +
> > +#define bte_zero_notify __get_cpu_var(bte_zero_notify)
> > +
> > +static inline struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
> > {
> > nodepda_t *tmp_nodepda;
> >
> > @@ -132,7 +140,6 @@
> > if (bte == NULL) {
> > continue;
> > }
> > -
> > if (spin_trylock(&bte->spinlock)) {
> > if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
> > (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
> > @@ -157,7 +164,7 @@
> > }
> > } while (1);
> >
> > - if (notification == NULL) {
> > + if (notification == NULL || (mode & BTE_NOTIFY_AND_GET_POINTER)) {
> > /* User does not want to be notified. */
> > bte->most_rcnt_na = &bte->notify;
> > } else {
> > @@ -192,6 +199,8 @@
> >
> > itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
> >
> > + if (mode & BTE_NOTIFY_AND_GET_POINTER)
> > + *(u64 volatile **)(notification) = &bte->notify;
> > spin_unlock_irqrestore(&bte->spinlock, irq_flags);
> >
> > if (notification != NULL) {
> > @@ -449,5 +458,31 @@
> > mynodepda->bte_if[i].cleanup_active = 0;
> > mynodepda->bte_if[i].bh_error = 0;
> > }
> > +}
> > +
> > +static int bte_check_bzero(void)
> > +{
> > + return *bte_zero_notify != BTE_WORD_BUSY;
> > +}
> > +
> > +static int bte_start_bzero(void *p, unsigned long len)
> > +{
> > + /* Check limitations.
> > + 1. System must be running (weird things happen during bootup)
> > + 2. Size >64KB. Smaller requests cause too much bte traffic
> > + */
> > + if (len >= BTE_MAX_XFER || len < 60000 || system_state != SYSTEM_RUNNING)
> > + return EINVAL;
> > +
> > + return bte_zero(ia64_tpa(p), len, BTE_NOTIFY_AND_GET_POINTER, &bte_zero_notify);
> > +}
> > +
> > +static struct zero_driver bte_bzero = {
> > + .start = bte_start_bzero,
> > + .check = bte_check_bzero,
> > + .rate = 500000000 /* 500 MB /sec */
> > +};
> >
> > +void sn_bte_bzero_init(void) {
> > + register_zero_driver(&bte_bzero);
> > }
> > Index: linux-2.6.9/arch/ia64/sn/kernel/setup.c
> > ===================================================================
> > --- linux-2.6.9.orig/arch/ia64/sn/kernel/setup.c 2004-12-17 14:40:10.000000000 -0800
> > +++ linux-2.6.9/arch/ia64/sn/kernel/setup.c 2004-12-21 11:02:35.000000000 -0800
> > @@ -243,6 +243,7 @@
> > int pxm;
> > int major = sn_sal_rev_major(), minor = sn_sal_rev_minor();
> > extern void sn_cpu_init(void);
> > + extern void sn_bte_bzero_init(void);
> >
> > /*
> > * If the generic code has enabled vga console support - lets
> > @@ -333,6 +334,7 @@
> > screen_info = sn_screen_info;
> >
> > sn_timer_init();
> > + sn_bte_bzero_init();
> > }
> >
> > /**
> > Index: linux-2.6.9/include/asm-ia64/sn/bte.h
> > ===================================================================
> > --- linux-2.6.9.orig/include/asm-ia64/sn/bte.h 2004-12-17 14:40:16.000000000 -0800
> > +++ linux-2.6.9/include/asm-ia64/sn/bte.h 2004-12-21 11:02:35.000000000 -0800
> > @@ -48,6 +48,8 @@
> > #define BTE_ZERO_FILL (BTE_NOTIFY | IBCT_ZFIL_MODE)
> > /* Use a reserved bit to let the caller specify a wait for any BTE */
> > #define BTE_WACQUIRE (0x4000)
> > +/* Return the pointer to the notification cacheline to the user */
> > +#define BTE_NOTIFY_AND_GET_POINTER (0x8000)
> > /* Use the BTE on the node with the destination memory */
> > #define BTE_USE_DEST (BTE_WACQUIRE << 1)
> > /* Use any available BTE interface on any node for the transfer */
>
next prev parent reply other threads:[~2004-12-22 19:58 UTC|newest]
Thread overview: 87+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <B8E391BBE9FE384DAA4C5C003888BE6F02900FBD@scsmsx401.amr.corp.intel.com>
[not found] ` <41C20E3E.3070209@yahoo.com.au>
2004-12-21 19:55 ` Increase page fault rate by prezeroing V1 [0/3]: Overview Christoph Lameter
2004-12-21 19:56 ` Increase page fault rate by prezeroing V1 [1/3]: Introduce __GFP_ZERO Christoph Lameter
2004-12-21 19:57 ` Increase page fault rate by prezeroing V1 [2/3]: zeroing and scrubd Christoph Lameter
2005-01-01 2:22 ` Nick Piggin
2005-01-01 2:55 ` pmarques
2004-12-21 19:57 ` Increase page fault rate by prezeroing V1 [3/3]: Altix SN2 BTE Zeroing Christoph Lameter
2004-12-22 12:46 ` Robin Holt
2004-12-22 19:56 ` Christoph Lameter [this message]
2004-12-23 19:29 ` Prezeroing V2 [0/3]: Why and When it works Christoph Lameter
2004-12-23 19:33 ` Prezeroing V2 [1/4]: __GFP_ZERO / clear_page() removal Christoph Lameter
2004-12-23 19:33 ` Prezeroing V2 [2/4]: add second parameter to clear_page() for all arches Christoph Lameter
2004-12-24 8:33 ` Pavel Machek
2004-12-24 16:18 ` Christoph Lameter
2004-12-24 16:27 ` Pavel Machek
2004-12-24 17:02 ` David S. Miller
2004-12-24 17:05 ` David S. Miller
2004-12-27 22:48 ` David S. Miller
2005-01-03 17:52 ` Christoph Lameter
2005-01-01 10:24 ` Geert Uytterhoeven
2005-01-04 23:12 ` Prezeroing V3 [0/4]: Discussion and i386 performance tests Christoph Lameter
2005-01-04 23:13 ` Prezeroing V3 [1/4]: Allow request for zeroed memory Christoph Lameter
2005-01-04 23:45 ` Dave Hansen
2005-01-05 1:16 ` Christoph Lameter
2005-01-05 1:26 ` Linus Torvalds
2005-01-05 23:11 ` Christoph Lameter
2005-01-05 0:34 ` Linus Torvalds
2005-01-05 0:47 ` Andrew Morton
2005-01-05 1:15 ` Christoph Lameter
2005-01-08 21:12 ` Hugh Dickins
2005-01-08 21:56 ` David S. Miller
2005-01-21 20:09 ` alloc_zeroed_user_highpage to fix the clear_user_highpage issue Christoph Lameter
2005-02-09 9:58 ` [Patch] Fix oops in alloc_zeroed_user_highpage() when page is NULL Michael Ellerman
2005-02-10 0:38 ` Christoph Lameter
2005-01-21 20:12 ` Extend clear_page by an order parameter Christoph Lameter
2005-01-21 22:29 ` Paul Mackerras
2005-01-21 23:48 ` Christoph Lameter
2005-01-22 0:35 ` Paul Mackerras
2005-01-22 0:43 ` Andrew Morton
2005-01-22 1:08 ` Paul Mackerras
2005-01-22 1:20 ` Roman Zippel
2005-01-22 1:25 ` Paul Mackerras
2005-01-22 1:54 ` Christoph Lameter
2005-01-22 2:53 ` Paul Mackerras
2005-01-23 7:45 ` Andrew Morton
2005-01-24 16:37 ` Christoph Lameter
2005-01-24 20:23 ` David S. Miller
2005-01-24 20:33 ` Christoph Lameter
2005-01-21 20:15 ` A scrub daemon (prezeroing) Christoph Lameter
2005-01-10 17:16 ` Prezeroing V3 [1/4]: Allow request for zeroed memory Christoph Lameter
2005-01-10 18:13 ` Linus Torvalds
2005-01-10 20:17 ` Christoph Lameter
2005-01-10 23:53 ` Prezeroing V4 [0/4]: Overview Christoph Lameter
2005-01-10 23:54 ` Prezeroing V4 [1/4]: Arch specific page zeroing during page fault Christoph Lameter
2005-01-11 0:41 ` Chris Wright
2005-01-11 0:46 ` Christoph Lameter
2005-01-11 0:49 ` Chris Wright
2005-01-10 23:55 ` Prezeroing V4 [2/4]: Zeroing implementation Christoph Lameter
2005-01-10 23:55 ` Prezeroing V4 [3/4]: Altix SN2 BTE zero driver Christoph Lameter
2005-01-10 23:56 ` Prezeroing V4 [4/4]: Extend clear_page to take an order parameter Christoph Lameter
2005-01-04 23:14 ` Prezeroing V3 [2/4]: Extension of " Christoph Lameter
2005-01-05 23:25 ` Christoph Lameter
2005-01-06 13:52 ` Andi Kleen
2005-01-06 17:47 ` Christoph Lameter
2005-01-04 23:15 ` Prezeroing V3 [3/4]: Page zeroing through kscrubd Christoph Lameter
2005-01-04 23:16 ` Prezeroing V3 [4/4]: Driver for hardware zeroing on Altix Christoph Lameter
2005-01-05 2:16 ` Andi Kleen
2005-01-05 16:24 ` Christoph Lameter
2004-12-23 19:34 ` Prezeroing V2 [3/4]: Add support for ZEROED and NOT_ZEROED free maps Christoph Lameter
2004-12-23 19:35 ` Prezeroing V2 [4/4]: Hardware Zeroing through SGI BTE Christoph Lameter
2004-12-23 20:08 ` Prezeroing V2 [1/4]: __GFP_ZERO / clear_page() removal Brian Gerst
2004-12-24 16:24 ` Christoph Lameter
2004-12-23 19:49 ` Prezeroing V2 [0/3]: Why and When it works Arjan van de Ven
2004-12-23 20:57 ` Matt Mackall
2004-12-23 21:01 ` Paul Mackerras
2004-12-23 21:11 ` Paul Mackerras
2004-12-23 21:37 ` Andrew Morton
2004-12-23 23:00 ` Paul Mackerras
2004-12-23 21:48 ` Linus Torvalds
2004-12-23 22:34 ` Zwane Mwaikambo
2004-12-24 9:14 ` Arjan van de Ven
2004-12-24 18:21 ` Linus Torvalds
2004-12-24 18:57 ` Arjan van de Ven
2004-12-27 22:50 ` David S. Miller
2004-12-28 11:53 ` Marcelo Tosatti
2004-12-24 16:17 ` Christoph Lameter
2004-12-24 18:31 ` Increase page fault rate by prezeroing V1 [0/3]: Overview Andrea Arcangeli
2005-01-03 17:54 ` Christoph Lameter
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=Pine.LNX.4.58.0412221154340.9186@schroedinger.engr.sgi.com \
--to=clameter@sgi.com \
--cc=agl@us.ibm.com \
--cc=holt@sgi.com \
--cc=linux-ia64@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@vger.kernel.org \
--cc=nickpiggin@yahoo.com.au \
--cc=tony.luck@intel.com \
--cc=torvalds@osdl.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).