2.6.7 backport request, spinoff idea
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 5 lines · Mikhail Ramendik
Quoted 2 lines · Con Kolivas
Actually I wonder if the swap token is responsible in 2.6.10-rc3
Quoted 7 lines
Quoted 2 lines
Try this
echo 0 >/proc/sys/vm/swap_token_timeoutWell, this makes things really better, although not as good as 2.6.8.1. There are kswapd cpu jumps and screen freezes, but they only last several seconds (the longest is in the very beginning, about 15-20 sec).
Thanks! This makes the system much more useable.
This is worth posting to lkml so I've cc'ed a few relevant people (there is already a thread about this there). Rik care to comment? I recall pointing out this test case a while back. Got any way to make it harder to trigger the swap token? It seems to affect normal workloads adversely considering how rare swap thrashing actually occurs.
Con
Quoted 12 lines · Con Kolivas
Quoted 10 lines · Mikhail Ramendik
Quoted 4 lines · Con Kolivas
Actually I wonder if the swap token is responsible in 2.6.10-rc3
Try this
echo 0 >/proc/sys/vm/swap_token_timeout
Quoted 6 lines
Quoted 3 lines
Well, this makes things really better, although not as good as 2.6.8.1. There are kswapd cpu jumps and screen freezes, but they only last several seconds (the longest is in the very beginning, about 15-20 sec).
Thanks! This makes the system much more useable.
This is worth posting to lkml so I've cc'ed a few relevant people (there is already a thread about this there). Rik care to comment? I recall pointing out this test case a while back. Got any way to make it harder to trigger the swap token? It seems to affect normal workloads adversely considering how rare swap thrashing actually occurs.
Rik tells me he has already posted a patch for this problem :D
Sorry for cc'ing the wrong email Rik. Do you have a simple link to the patch or can you send it on this thread please?
Cheers,
Con
Quoted 1 line · Con Kolivas
Sorry for cc'ing the wrong email Rik. Do you have a simple link to the patch or can you send it on this thread please?
Here it is. The patch should have the effect of reducing the swap token timeout to the minimum value at times when there is little to no memory pressure, while allowing it to grow in times where it is needed.
Date: Sat, 11 Dec 2004 14:24:15 -0500 (EST)
From: Rik van Riel <[email protected]>
To: Marcelo Tosatti <[email protected]>
Cc: [email protected]
Subject: PATCH: automatic tuning of swap token timeout
At Marcelo's request. I made this patch a while ago so I'm not sure if it will still apply to the recent kernel, but it should give you an idea of what I tried to achieve.
The idea is to keep the swap token timeout short on a system with mostly small tasks, even if there is one big hog running in the background. The big thrashing program should not hold the swap token for unfair amounts of time.
The swap token timeout should be the minimum required to keep most of the processes in the system from thrashing, while keeping some amount of fairness.
This patch is untested. Have fun.
===== mm/thrash.c 1.2 vs edited =====
mm/thrash.c +44 -3
@@ -19,10 +19,44 @@struct mm_struct * swap_token_mm = &init_mm;#define SWAP_TOKEN_CHECK_INTERVAL (HZ * 2) -#define SWAP_TOKEN_TIMEOUT (HZ * 300) +#define MIN_SWAP_TOKEN_TIMEOUT (2 * SWAP_TOKEN_CHECK_INTERVAL) +#define SWAP_TOKEN_TIMEOUT (HZ * 30) +#define MAX_SWAP_TOKEN_TIMEOUT (HZ * 300)unsigned long swap_token_default_timeout =SWAP_TOKEN_TIMEOUT ;/* +* We count how often the swap token times out, and how often the +* swap token hold time is long enough for processes to regain their +* working set and make progress. +* +* The goal is to have processes in the system make progress, with the +* lowest possible latency. If the token times out too often, processes +* are not making progress and the timeout needs to be increased. If +* processes are making progress, we can decrease the timeout and improve +* system latency. +*/ +#define SWAP_TOKEN_RECALC 32 +#define SWAP_TOKEN_TOO_LONG 1 +static int swap_token_timed_out; +static int swap_token_enough_rss; + +static void recalculate_swap_token_timeout (void ) +{ +unsigned long delta = (swap_token_default_timeout /4 ) +1 ; +if (swap_token_timed_out >SWAP_TOKEN_TOO_LONG ) { + swap_token_default_timeout += delta; +if (swap_token_default_timeout >MAX_SWAP_TOKEN_TIMEOUT ) + swap_token_default_timeout =MAX_SWAP_TOKEN_TIMEOUT ; + }else { + swap_token_default_timeout -= delta; +if (swap_token_default_timeout <MIN_SWAP_TOKEN_TIMEOUT ) + swap_token_default_timeout =MIN_SWAP_TOKEN_TIMEOUT ; + } + swap_token_enough_rss /=2 ; + swap_token_timed_out /=2 ; +} + +/* * Take the token away if the process had no page faults * in the last interval, or if it has held the token for * too long.
@@ -32,11 +66,18 @@static int should_release_swap_token (struct mm_struct *mm) {int ret =0 ; -if (!mm->recent_pagein) +if (!mm->recent_pagein) { + swap_token_enough_rss++; ret =SWAP_TOKEN_ENOUGH_RSS ; -else if (time_after (jiffies, swap_token_timeout)) + }else if (time_after (jiffies, swap_token_timeout)) { + swap_token_timed_out++; ret =SWAP_TOKEN_TIMED_OUT ; + } mm->recent_pagein =0 ; + +if (swap_token_timed_out + swap_token_enough_rss >SWAP_TOKEN_RECALC ) +recalculate_swap_token_timeout (); +return ret; }