Skip to contentSkip to conversation

vm-thrashing-control-tuning

PATCH In review Hideo AOKI 3 participants 4 messages Oct 5, 2004 – Oct 6, 2004

Subsystems (3)
lore
Share & tools
lore

Hideo AOKI

author
#0 lore raw

PATCH 1/1 vm-thrashing-control-tuning

Hello,

Token based thrashing control has a good effect in thrashing situation.
In current implementation, token timeout is a fixed value as SWAP_TOKEN_TIMEOUT. I think that administrators can improve thrashing behavior if the value can be changed.

This patch adds "swap_token_timeout" parameter in /proc/sys/vm.
The parameter means expired time of token. Unit of the value is HZ, and the default value is the same as current SWAP_TOKEN_TIMEOUT (i.e. HZ * 300). The patch can be applied to both 2.6.9-rc2 and 2.6.9-rc3.

I tested the patch on an IA-32 4 way machine which has 4GB memory. Based kernel was 2.6.9-rc1. I tested five swap_token_time_out values. I created 256 workload generation processes on the machine. Each process generated same workload: repeating random file writing and memory access (Memory region is 18MB). I ran the processes for one hour and calculated write throughput of workload processes. The result was following.

swap_token_time_out   Write throughput [MB/s]
-------------------   ----------------------
30,000,000            6.71
3,000,000            8.26
  300,000 (DEFAULT)  7.90
   30,000            8.16
    3,000            7.43

As you can see, it may be possible to improve application performance
according to tune swap_token_time_out. Additionally, I think it is better to decrease default value. One reason is that the other values gained good performance. The other is that behavior of kernel may be unstable if swap_token_time_out is too long.

I am exploring tuning policy.

Any comments or suggestions?

Best regards,

Hideo AOKI

Systems Development Laboratory, Hitachi, Ltd.

----

include/linux/swap.h 2004-09-28 15:12:33.000000000 +0900 renamed from include/linux/swap.h 2004-09-15 12:20:40.000000000 +0900 +0 -0
diff -uprN linux-2.6.9-rc2/include/linux/swap.h linux-2.6.9-rc2-tbtc_tune/include/linux/swap.h
@@ -230,6 +230,7 @@ extern spinlock_t swaplock;
 

/* linux/mm/thrash.c */
extern struct mm_struct * swap_token_mm;
+extern unsigned long swap_token_default_timeout;
extern void grab_swap_token(void);
extern void __put_swap_token(struct mm_struct *);

include/linux/sysctl.h 2004-09-28 15:12:33.000000000 +0900 renamed from include/linux/sysctl.h 2004-09-15 12:20:40.000000000 +0900 +1 -0
diff -uprN linux-2.6.9-rc2/include/linux/sysctl.h linux-2.6.9-rc2-tbtc_tune/include/linux/sysctl.h
@@ -167,6 +167,7 @@ enum
   VM_HUGETLB_GROUP=25,    /* permitted hugetlb group */
   VM_VFS_CACHE_PRESSURE=26, /* dcache/icache reclaim pressure */
   VM_LEGACY_VA_LAYOUT=27, /* legacy/compatibility virtual address space layout */
+    VM_SWAP_TOKEN_TIMEOUT=28 /* default time for token time out */

};

kernel/sysctl.c 2004-09-28 15:12:33.000000000 +0900 renamed from kernel/sysctl.c 2004-09-15 12:20:41.000000000 +0900 +0 -0
diff -uprN linux-2.6.9-rc2/kernel/sysctl.c linux-2.6.9-rc2-tbtc_tune/kernel/sysctl.c
@@ -800,6 +800,15 @@ static ctl_table vm_table[] = {
       .extra1        = &zero,
   },

#endif
+ {

+        .ctl_name    = VM_SWAP_TOKEN_TIMEOUT,
+        .procname    = "swap_token_timeout",
+        .data        = &swap_token_default_timeout,
+        .maxlen        = sizeof(swap_token_default_timeout),
+        .mode        = 0644,
+        .proc_handler    = &proc_dointvec,
+        .strategy    = &sysctl_intvec,

+ },
{ .ctl_name = 0 }
};

mm/thrash.c 2004-09-28 15:12:33.000000000 +0900 renamed from mm/thrash.c 2004-09-15 12:20:42.000000000 +0900 +0 -0
diff -uprN linux-2.6.9-rc2/mm/thrash.c linux-2.6.9-rc2-tbtc_tune/mm/thrash.c
@@ -20,6 +20,8 @@ struct mm_struct * swap_token_mm = &init
 

#define SWAP_TOKEN_CHECK_INTERVAL (HZ * 2)
#define SWAP_TOKEN_TIMEOUT (HZ * 300)
+unsigned long swap_token_default_timeout = SWAP_TOKEN_TIMEOUT; +

/*
* Take the token away if the process had no page faults @@ -75,10 +77,10 @@ void grab_swap_token(void)

       if ((reason = should_release_swap_token(mm))) {
           unsigned long eligible = jiffies;
           if (reason == SWAP_TOKEN_TIMED_OUT) {
-                eligible += SWAP_TOKEN_TIMEOUT;
+                eligible += swap_token_default_timeout;
           }
           mm->swap_token_time = eligible;
-            swap_token_timeout = jiffies + SWAP_TOKEN_TIMEOUT;
+            swap_token_timeout = jiffies + swap_token_default_timeout;
           swap_token_mm = current->mm;
       }
       spin_unlock(&swap_token_lock);
#1 lore raw
Quoted 3 lines · Hideo AOKI

I am exploring tuning policy.

Any comments or suggestions?

While I believe that a self tuning timeout might be better in the long run, this tunable will certainly help tune policy.

Besides, even with a self tuning timeout we'll want to have this tunable visible in /proc so we can debug things by seeing what value the kernel set the tunable to ;)

Signature · 3 lines
"Debugging is twice as hard as writing the code in the first place.
Therefore, if you write the code as cleverly as possible, you are,
by definition, not smart enough to debug it." - Brian W. Kernighan
author
#2 lore raw
Quoted 2 lines · Rik van Riel

While I believe that a self tuning timeout might be better in the long run, this tunable will certainly help tune policy.

Thank you for your comments.

I agree with you. The best solution is a self tuning timeout.
Using this swap_token_timeout parameter, I would like to get a clue to self tuning.

Best regards,

Hideo AOKI

Systems Development Laboratory, Hitachi, Ltd.

#3 lore raw
In reply to Hideo AOKI
Quoted 2 lines · Hideo AOKI

User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.1; ja-JP; rv:1.4) Gecko/20030624 Netscape/7.1 (ax)

uh-oh.

Quoted 1 line

This patch adds "swap_token_timeout" parameter in /proc/sys/vm.

The patch is hopelessly whitespace-mangled. Please use attachments.