NFS remove readdir plus limit
We will now use readdir plus even on directories that are very large.
Signed-off-by: Bryan Schumaker <[email protected]>
---
diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index 7d2d6c7..b2e12bc 100644
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -234,9 +234,6 @@ nfs_init_locked(struct inode *inode, void *opaque)
return 0;
}
-/* Don't use READDIRPLUS on directories that we believe are too large */
-#define NFS_LIMIT_READDIRPLUS (8*PAGE_SIZE)
-
/*
* This is our front-end to iget that looks up inodes by file handle
* instead of inode number.
@@ -291,8 +288,7 @@ nfs_fhget(struct super_block *sb, struct nfs_fh *fh, struct nfs_fattr *fattr)
} else if (S_ISDIR(inode->i_mode)) {
inode->i_op = NFS_SB(sb)->nfs_client->rpc_ops->dir_inode_ops;
inode->i_fop = &nfs_dir_operations;
- if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS)
- && fattr->size <= NFS_LIMIT_READDIRPLUS)
+ if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS))
set_bit(NFS_INO_ADVISE_RDPLUS, &NFS_I(inode)->flags);
/* Deal with crossing mountpoints */
if ((fattr->valid & NFS_ATTR_FATTR_FSID)
Thanks for the advice. I did some testing between my machine (server) and a virtual machine (client). The command I ran was ls -l --color=none on a directory with 10,000 files. With the directory cap, nfsstat had the following output for both a stock kernel and a kernel with these patches applied
Stock Kernel
---------------------
calls: 10101
getattr: 3
readdir: 89
lookup: 10002
My Kernel
----------------------
calls: 10169
getattr: 3
readdir: 157
lookup: 10002
Without the directory cap, I saw the following numbers
Trial 1
--------------------
calls: 1710
getattr: 1622
readdirplus: 79
lookup: 2
Trial 2
--------------------
calls: 1233
getattr: 1145
readdirplus: 79
lookup: 2
Trial 3
--------------------
calls: 217
getattr: 129
readdirplus: 79
lookup: 2
In each of these cases the number of lookups has dropped from 10,002 to 2. The number of total calls has dropped significantly as well. I suspect that the change in getattrs is caused by a race between items hitting the cache and the same items being used.
Let me know if I should run other tests.
Bryan
On 09/07/2010 04:33 PM, Chuck Lever wrote:
> Hi Bryan-
>
> On Sep 7, 2010, at 4:03 PM, Bryan Schumaker wrote:
>
>> NFS remove readdir plus limit
>>
>> We will now use readdir plus even on directories that are very large.
>
> READDIRPLUS operations on some servers may be quite expensive, since the server usually treats directories as byte streams, and can be read sequentially; but inode attributes are read by random disk seeks. So assembling a READDIRPLUS result on a large directory that isn't in the server's cache might be an awful lot of work on a busy server.
>
> On large directories, there isn't much proven benefit to having all the dcache entries on hand on the client. It can even hurt performance by pushing more useful entries out of the cache.
>
> If we really want to take the directory size cap off, that seems like it could be a far-reaching change. You should at least use the patch description to provide thorough rationale. Even some benchmark results, with especially slow servers and networks, and small clients, would be nice.
>
>> Signed-off-by: Bryan Schumaker <[email protected]>
>> ---
>> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
>> index 7d2d6c7..b2e12bc 100644
>> --- a/fs/nfs/inode.c
>> +++ b/fs/nfs/inode.c
>> @@ -234,9 +234,6 @@ nfs_init_locked(struct inode *inode, void *opaque)
>> return 0;
>> }
>>
>> -/* Don't use READDIRPLUS on directories that we believe are too large */
>> -#define NFS_LIMIT_READDIRPLUS (8*PAGE_SIZE)
>> -
>> /*
>> * This is our front-end to iget that looks up inodes by file handle
>> * instead of inode number.
>> @@ -291,8 +288,7 @@ nfs_fhget(struct super_block *sb, struct nfs_fh *fh, struct nfs_fattr *fattr)
>> } else if (S_ISDIR(inode->i_mode)) {
>> inode->i_op = NFS_SB(sb)->nfs_client->rpc_ops->dir_inode_ops;
>> inode->i_fop = &nfs_dir_operations;
>> - if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS)
>> - && fattr->size <= NFS_LIMIT_READDIRPLUS)
>> + if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS))
>> set_bit(NFS_INO_ADVISE_RDPLUS, &NFS_I(inode)->flags);
>> /* Deal with crossing mountpoints */
>> if ((fattr->valid & NFS_ATTR_FATTR_FSID)
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
Hi Bryan-
On Sep 7, 2010, at 4:03 PM, Bryan Schumaker wrote:
> NFS remove readdir plus limit
>
> We will now use readdir plus even on directories that are very large.
READDIRPLUS operations on some servers may be quite expensive, since the server usually treats directories as byte streams, and can be read sequentially; but inode attributes are read by random disk seeks. So assembling a READDIRPLUS result on a large directory that isn't in the server's cache might be an awful lot of work on a busy server.
On large directories, there isn't much proven benefit to having all the dcache entries on hand on the client. It can even hurt performance by pushing more useful entries out of the cache.
If we really want to take the directory size cap off, that seems like it could be a far-reaching change. You should at least use the patch description to provide thorough rationale. Even some benchmark results, with especially slow servers and networks, and small clients, would be nice.
> Signed-off-by: Bryan Schumaker <[email protected]>
> ---
> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> index 7d2d6c7..b2e12bc 100644
> --- a/fs/nfs/inode.c
> +++ b/fs/nfs/inode.c
> @@ -234,9 +234,6 @@ nfs_init_locked(struct inode *inode, void *opaque)
> return 0;
> }
>
> -/* Don't use READDIRPLUS on directories that we believe are too large */
> -#define NFS_LIMIT_READDIRPLUS (8*PAGE_SIZE)
> -
> /*
> * This is our front-end to iget that looks up inodes by file handle
> * instead of inode number.
> @@ -291,8 +288,7 @@ nfs_fhget(struct super_block *sb, struct nfs_fh *fh, struct nfs_fattr *fattr)
> } else if (S_ISDIR(inode->i_mode)) {
> inode->i_op = NFS_SB(sb)->nfs_client->rpc_ops->dir_inode_ops;
> inode->i_fop = &nfs_dir_operations;
> - if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS)
> - && fattr->size <= NFS_LIMIT_READDIRPLUS)
> + if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS))
> set_bit(NFS_INO_ADVISE_RDPLUS, &NFS_I(inode)->flags);
> /* Deal with crossing mountpoints */
> if ((fattr->valid & NFS_ATTR_FATTR_FSID)
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
chuck[dot]lever[at]oracle[dot]com
On Sep 8, 2010, at 9:10 AM, Bryan Schumaker wrote:
> Thanks for the advice. I did some testing between my machine (server) and a virtual machine (client). The command I ran was ls -l --color=none on a directory with 10,000 files. With the directory cap, nfsstat had the following output for both a stock kernel and a kernel with these patches applied
>
> Stock Kernel
> ---------------------
> calls: 10101
> getattr: 3
> readdir: 89
> lookup: 10002
>
>
> My Kernel
> ----------------------
> calls: 10169
> getattr: 3
> readdir: 157
> lookup: 10002
>
>
> Without the directory cap, I saw the following numbers
>
> Trial 1
> --------------------
> calls: 1710
> getattr: 1622
> readdirplus: 79
> lookup: 2
>
> Trial 2
> --------------------
> calls: 1233
> getattr: 1145
> readdirplus: 79
> lookup: 2
>
> Trial 3
> --------------------
> calls: 217
> getattr: 129
> readdirplus: 79
> lookup: 2
>
>
> In each of these cases the number of lookups has dropped from 10,002 to 2. The number of total calls has dropped significantly as well. I suspect that the change in getattrs is caused by a race between items hitting the cache and the same items being used.
>
> Let me know if I should run other tests.
"ls -l" is, of course, an important test. These results are as successful as one could hope, and demonstrate that there will be some benefit if the cap is, at least, raised, as I'm sure the average number of entries in directories is increasing over time. My worry is what happens if the cap is left unlimited. To test this, here are a few ideas:
How about trying a directory with a million entries, or ten million? What happens to system memory utilization over time when your client has 128MB of RAM or less, or has several other processes that are hogging memory? What is the impact of "ls -l" on other processes that are using the dcache? Did you measure how expensive (CPU and disk operations) a large readdirplus request is on the server, compared to a similar stream of READDIR and GETATTR or LOOKUP requests? What is the average latency of readdirplus requests as they grow larger (this would test the other patch that increases the maximum size of readdirplus replies).
What happens if the workload is not a long sequence of stat(2) calls but instead the application is opening files in a very large directory and reading or writing them (similar to a web server workload)? After the directory is enumerated on the client, it will continue to issue OPEN, LOOKUP, or GETATTR calls anyway. Is the extra attribute data returned by the READDIRPLUS of any use in this case? Is there a difference in behavior or performance between NFSv3 and NFSv4?
What is the network and server overhead when the target directory changes frequently? In other words, is using READDIRPLUS to return this information amount to more or fewer bytes on the network, or disk operations on the server? If files are added or removed from the directory often, will the client be doing more work or less to refresh its caches?
Thanks for giving it a shot.
> Bryan
>
> On 09/07/2010 04:33 PM, Chuck Lever wrote:
>> Hi Bryan-
>>
>> On Sep 7, 2010, at 4:03 PM, Bryan Schumaker wrote:
>>
>>> NFS remove readdir plus limit
>>>
>>> We will now use readdir plus even on directories that are very large.
>>
>> READDIRPLUS operations on some servers may be quite expensive, since the server usually treats directories as byte streams, and can be read sequentially; but inode attributes are read by random disk seeks. So assembling a READDIRPLUS result on a large directory that isn't in the server's cache might be an awful lot of work on a busy server.
>>
>> On large directories, there isn't much proven benefit to having all the dcache entries on hand on the client. It can even hurt performance by pushing more useful entries out of the cache.
>>
>> If we really want to take the directory size cap off, that seems like it could be a far-reaching change. You should at least use the patch description to provide thorough rationale. Even some benchmark results, with especially slow servers and networks, and small clients, would be nice.
>>
>>> Signed-off-by: Bryan Schumaker <[email protected]>
>>> ---
>>> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
>>> index 7d2d6c7..b2e12bc 100644
>>> --- a/fs/nfs/inode.c
>>> +++ b/fs/nfs/inode.c
>>> @@ -234,9 +234,6 @@ nfs_init_locked(struct inode *inode, void *opaque)
>>> return 0;
>>> }
>>>
>>> -/* Don't use READDIRPLUS on directories that we believe are too large */
>>> -#define NFS_LIMIT_READDIRPLUS (8*PAGE_SIZE)
>>> -
>>> /*
>>> * This is our front-end to iget that looks up inodes by file handle
>>> * instead of inode number.
>>> @@ -291,8 +288,7 @@ nfs_fhget(struct super_block *sb, struct nfs_fh *fh, struct nfs_fattr *fattr)
>>> } else if (S_ISDIR(inode->i_mode)) {
>>> inode->i_op = NFS_SB(sb)->nfs_client->rpc_ops->dir_inode_ops;
>>> inode->i_fop = &nfs_dir_operations;
>>> - if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS)
>>> - && fattr->size <= NFS_LIMIT_READDIRPLUS)
>>> + if (nfs_server_capable(inode, NFS_CAP_READDIRPLUS))
>>> set_bit(NFS_INO_ADVISE_RDPLUS, &NFS_I(inode)->flags);
>>> /* Deal with crossing mountpoints */
>>> if ((fattr->valid & NFS_ATTR_FATTR_FSID)
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
>>> the body of a message to [email protected]
>>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>
>
--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com