make nfsd_drc_max_mem configurable

PATCH In review Christoph Martin 2 participants 4 messages Jun 17, 2015 – Aug 20, 2015

Subsystems (2)

NFS client NFS server

lore

Share & tools

lore

Christoph Martin

author

2015-06-17 12:59 UTC #0 lore raw

PATCH 1/1 make nfsd_drc_max_mem configurable

Dear Andy, dear Bruce,

(sorry for the recent to you, I now cc'd linux-nfs)

I have attached a patch to nfssvc.diff to make the size of the drc nfsd cache configurable.

In the last month we were stumbling twice over the problem that the NFS4.1 session cache was to small.

The first time we wanted to setup a NFS Server for our HPC cluster. We were wondering why we were only able to mount the filesystem on 380 of our ~700 nodes. It took us a long time to find out that it was the limit of the NFS4.1 session cache. Since this machine had 12G Ram, the kernel reserved 12M for the cache, which results in 384 slots a 32k:

echo $(((12582912>>10)/32))
384

We patched the kernel redhat 7 kernel to change NFSD_DRC_SIZE_SHIFT to from 10 to 7 to fix this problem.

The second time we installed a small Debian VM with 1G ram to act as a NFS4 referral server for the home and group directories on our campus. Since the server does only NFS referrals it does not really need more memory than the 1G. But it could only server about 30 clients with this limitation of the session cache.

I think it would be a good idea to have the amount of memory configurable in nfsd. So I wrote this small patch to make drc_size configurable while loading the kernel nfsd module.

The patch uses the old value computed from NFSD_DRC_SIZE_SHIFT as the lower limit. If drc_size as a parameter for then nfsd is higher than a 1/1000 of the RAM, this value will be used.

One might consider to make NFSD_DRC_SIZE_SHIFT even higher to use less memory for situations where it is not needed. I did not implement an upper limit, but it might be important.

Please consider to include this patch into the nfsd code.

Yours
Christoph Martin

--- linux-source-3.16/fs/nfsd/nfssvc.c 2015-03-30 12:09:09.000000000 +0200
+++ linux-source-3.16.nfsd/fs/nfsd/nfssvc.c 2015-06-17
09:28:37.880443867 +0200
@@ -359,11 +359,19 @@ void nfsd_reset_versions(void)
* For now this is a #defined shift which could be under admin control
* in the future.
*/
+
+static ulong drc_size = 0;
+module_param(drc_size, ulong, 0444); +MODULE_PARM_DESC(drc_size,
+ "size of NFSv4.1 DRC cache memory (default and minimum:
free_buffer_size >> 10)");
+
static void set_max_drc(void)
{
#define NFSD_DRC_SIZE_SHIFT 10
- nfsd_drc_max_mem = (nr_free_buffer_pages()
- >> NFSD_DRC_SIZE_SHIFT) * PAGE_SIZE;
+ nfsd_drc_max_mem = max(drc_size,
+ (nr_free_buffer_pages()
+ >> NFSD_DRC_SIZE_SHIFT) * PAGE_SIZE);
+ drc_size = nfsd_drc_max_mem;
nfsd_drc_mem_used = 0;
spin_lock_init(&nfsd_drc_lock);
dprintk("%s nfsd_drc_max_mem %lu \n", __func__, nfsd_drc_max_mem);

J. Bruce Fields

2015-06-18 16:16 UTC #1 lore raw

Quoted 9 lines · Christoph Martin

Dear Andy, dear Bruce,
(sorry for the recent to you, I now cc'd linux-nfs)
I have attached a patch to nfssvc.diff to make the size of the drc nfsd cache configurable.
In the last month we were stumbling twice over the problem that the NFS4.1 session cache was to small.

Thanks for the careful investigation!

Quoted 8 lines

The first time we wanted to setup a NFS Server for our HPC cluster. We were wondering why we were only able to mount the filesystem on 380 of our ~700 nodes. It took us a long time to find out that it was the limit of the NFS4.1 session cache. Since this machine had 12G Ram, the kernel reserved 12M for the cache, which results in 384 slots a 32k:
echo $(((12582912>>10)/32))
384

So each client is using 32k?

Might be interesting to take a look at the CREATE_SESSION call and reply in wireshark (especially the values of maxresponsesize_cached and maxrequests)--there might also be defaults there that need tweaking.

Quoted 17 lines

We patched the kernel redhat 7 kernel to change NFSD_DRC_SIZE_SHIFT to from 10 to 7 to fix this problem.
The second time we installed a small Debian VM with 1G ram to act as a NFS4 referral server for the home and group directories on our campus. Since the server does only NFS referrals it does not really need more memory than the 1G. But it could only server about 30 clients with this limitation of the session cache.
I think it would be a good idea to have the amount of memory configurable in nfsd. So I wrote this small patch to make drc_size configurable while loading the kernel nfsd module.
The patch uses the old value computed from NFSD_DRC_SIZE_SHIFT as the lower limit. If drc_size as a parameter for then nfsd is higher than a 1/1000 of the RAM, this value will be used.

Quoted 5 lines

One might consider to make NFSD_DRC_SIZE_SHIFT even higher to use less memory for situations where it is not needed. I did not implement an upper limit, but it might be important.
Please consider to include this patch into the nfsd code.

Looks good, my one concern is that this covers only the size of the 4.1 session cache. We may need to add some more limits in the future and might not want to require separate configuration of each limit.

Maybe one or two more generic size parameters would be more useful? Like:

- Maximum memory to devote to knfsd
- Maximum memory to devote to a single client

--b.

Quoted 32 lines

Yours
Christoph Martin

fs/nfsd/nfssvc.c +0 -0

09:28:37.880443867 +0200

fs/nfsd/nfssvc.c +10 -2

@@ -359,11 +359,19 @@ void nfsd_reset_versions(void)
  * For now this is a #defined shift which could be under admin control
  * in the future.
  */
+
+static ulong drc_size = 0;
+module_param(drc_size, ulong, 0444);
+MODULE_PARM_DESC(drc_size,
+		 "size of NFSv4.1 DRC cache memory (default and minimum:
 free_buffer_size >> 10)");
+
 static void set_max_drc(void)
 {
 	#define NFSD_DRC_SIZE_SHIFT	10
-	nfsd_drc_max_mem = (nr_free_buffer_pages()
-					>> NFSD_DRC_SIZE_SHIFT) * PAGE_SIZE;
+	nfsd_drc_max_mem = max(drc_size,
+			       (nr_free_buffer_pages()
+				>> NFSD_DRC_SIZE_SHIFT) * PAGE_SIZE);
+	drc_size = nfsd_drc_max_mem;
 	nfsd_drc_mem_used = 0;
 	spin_lock_init(&nfsd_drc_lock);

dprintk("%s nfsd_drc_max_mem %lu \n", __func__, nfsd_drc_max_mem);

Christoph Martin

author

2015-07-06 12:59 UTC #2 lore raw

Hi Bruce,

Am 18.06.2015 um 18:16 schrieb J. Bruce Fields:

Quoted 11 lines

Quoted 9 lines
The first time we wanted to setup a NFS Server for our HPC cluster. We were wondering why we were only able to mount the filesystem on 380 of our ~700 nodes. It took us a long time to find out that it was the limit of the NFS4.1 session cache. Since this machine had 12G Ram, the kernel reserved 12M for the cache, which results in 384 slots a 32k:
echo $(((12582912>>10)/32))
384
So each client is using 32k?

#define NFSD_SLOT_CACHE_SIZE 2048
/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
#define NFSD_CACHE_SIZE_SLOTS_PER_SESSION 32
#define NFSD_MAX_MEM_PER_SESSION \
(NFSD_CACHE_SIZE_SLOTS_PER_SESSION * NFSD_SLOT_CACHE_SIZE)

So this would be 64k. Maybe I missed a factor of 2 somewhere. But the calculation above equals the experience in our tests.

Quoted 22 lines

Might be interesting to take a look at the CREATE_SESSION call and reply in wireshark (especially the values of maxresponsesize_cached and maxrequests)--there might also be defaults there that need tweaking.
Quoted 17 lines
We patched the kernel redhat 7 kernel to change NFSD_DRC_SIZE_SHIFT to from 10 to 7 to fix this problem.
The second time we installed a small Debian VM with 1G ram to act as a NFS4 referral server for the home and group directories on our campus. Since the server does only NFS referrals it does not really need more memory than the 1G. But it could only server about 30 clients with this limitation of the session cache.
I think it would be a good idea to have the amount of memory configurable in nfsd. So I wrote this small patch to make drc_size configurable while loading the kernel nfsd module.
The patch uses the old value computed from NFSD_DRC_SIZE_SHIFT as the lower limit. If drc_size as a parameter for then nfsd is higher than a 1/1000 of the RAM, this value will be used.

Quoted 7 lines

Quoted 5 lines
One might consider to make NFSD_DRC_SIZE_SHIFT even higher to use less memory for situations where it is not needed. I did not implement an upper limit, but it might be important.
Please consider to include this patch into the nfsd code.
Looks good,

As far as I understand the code now, it would even be possible to change the value of nfsd_drc_max_mem during runtime of nfsd, since the value is only used in nfsd4_get_drc_mem in nfs4state.c. I don't see that the limit is on the slab. It seems only to be on the local usage of the slab.

Quoted 10 lines

my one concern is that this covers only the size of the 4.1 session cache. We may need to add some more limits in the future and might not want to require separate configuration of each limit.
Maybe one or two more generic size parameters would be more useful? Like:
- Maximum memory to devote to knfsd
- Maximum memory to devote to a single client

We were discussing this and don't think that this is a good idea.

If you have only one limit per knfsd or client, you don't know how many memory you have to assign to the different memory slabs, like drc or others, because you can't know in advance if the nfs server is only used for say nfs3 or nfs4 or nfs4.1 or a mixture.

So if you think it is necessary to have a global limit, you also need tunabels for the distribution of the available memory to the different protocols.

I found the following calls to kmem_cache_create:

Quoted 7 lines

nfs4state.c:2635: openowner_slab = kmem_cache_create("nfsd4_openowners",
nfs4state.c:2639: lockowner_slab = kmem_cache_create("nfsd4_lockowners",
nfs4state.c:2643: file_slab = kmem_cache_create("nfsd4_files",
nfs4state.c:2647: stateid_slab = kmem_cache_create("nfsd4_stateids",
nfs4state.c:2651: deleg_slab = kmem_cache_create("nfsd4_delegations",
nfscache.c:168: drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),

At a first look, they all use different methods to administer this memory if at all.

Yours
Christoph

begin:vcard
fn:Christoph Martin
n:Martin;Christoph
email;internet:[email protected]
tel;work:+49-6131-3926337
tel;fax:+49-6131-3926407
tel;cell:+49-179-7952652
version:2.1
end:vcard

Christoph Martin

author

2015-08-20 09:40 UTC #3 lore raw

Hi Bruce

Quoted 11 lines

I found the following calls to kmem_cache_create:
Quoted 7 lines
nfs4state.c:2635: openowner_slab = kmem_cache_create("nfsd4_openowners",
nfs4state.c:2639: lockowner_slab = kmem_cache_create("nfsd4_lockowners",
nfs4state.c:2643: file_slab = kmem_cache_create("nfsd4_files",
nfs4state.c:2647: stateid_slab = kmem_cache_create("nfsd4_stateids",
nfs4state.c:2651: deleg_slab = kmem_cache_create("nfsd4_delegations",
nfscache.c:168: drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),

Quoted 3 lines

At a first look, they all use different methods to administer this memory if at all.

Another solution would be to simply remove the upper limit for the drc cache size.

What do you think? How do we solve this problem?

Christoph