CITI ASC status
From Linux NFS
University of Michigan/CITI NFSv4 ASC alliance
Status of October 2006
Task 1. pNFS Demonstration
Demonstration of pNFS with multiple back end methods (PVFS2 and File) including layout recall — LANL will replicate this demonstration at LANL working with CITI remotely
We updated the Linux pNFS client and server to the 2.6.17 kernel level, and are preparing to rebase again for 2.6.19.
We updated the pNFS code base to draft-ietf-nfsv4-minorversion1-05. Testing identified multiple bugs, which we fixed.
The linux client separates common NFS code from NFSv2/3/4 code by using version specific operations. We rewrote the Linux pNFS client to use its own set of version specfic operations. This provides a controlled interface to the pNFS code, and eases updating the code to new kernel versions.
Four client layout modules are in development.
- File layout driver (CITI, Network Appliance, and IBM Almaden).
- PVFS2 layout driver (CITI).
- Object layout driver (Panasas).
- Block layout driver (CITI).
To accommodate the requirements of the multiple layout drivers, we expanded the policy interface between the layout driver and generic pNFS client. This interface allows layout drivers to set the following policies:
- stripe size
- writeback cache gathering policies
- read and write threshold
- timing of layoutget invocation
- choice of pagecache or direct method for I/O
We are designing and coding a pNFS client layout cache to replace the current implementation, which supports only a single layout per inode.
We improved the interface to the underlying file system on the Linux pNFS server. The new interface is being used by the Panasas object layout server, the IBM GPFS server, and the PVFS2 server.
We are coding the pNFS layout management service and file system interfaces on the Linux pNFS server to do a better job of bookkeeping so that we can extend the layout recall implementation, which is limited to a single layout.
We have continued to develop the PVFS2 layout driver and PVFS2 support in the pNFS server. The layout driver I/O interface supports direct access, page cache access with NFSv4 readahead and writeback, and the O_DIRECT access method. In addition, PVFS2 now supports the pNFS file-based layout, which lets pNFS clients choose how they access the file system.
We demonstrated how pNFS can improve the overall write performance of parallel file systems by using direct, parallel I/O for large write requests and the NFSv4 storage protocol for small write requests. To switch between them, we added a write threshold to the layout driver. Write requests smaller than the threshold follow the slower NFSv4 data path. Write requests larger than the threshold follow the faster layout driver data path. D. Hildebrand, L. Ward, and P. Honeyman, "Large Files, Small Writes, and pNFS," in Proc. of the 20th ACM International Conf. on Supercomputing, Cairns, Australia, 2006.
We improved the performance and scalability of pNFS file-based access with parallel file systems. Our design, named Direct-pNFS, augmented the file-based architecture to enable file-based pNFS clients to bypass intermediate data servers and access heterogeneous data stores directly. Direct access is possible by ensuring file-based layouts match the data layout in the underlying file system and giving pNFS clients the tools to effectively interpret and utilize this information. Experiments with Direct-pNFS demonstrate I/O throughput that equals or outperforms the exported parallel file system across a range of workloads. D. Hildebrand and P. Honeyman, "Direct-pNFS: Simple, Transparent, and Versatile Access to Parallel File Systems," CITI Technical Report 06-8, October 2006.
We developed prototype implementations of pNFS operations:
- OP_LAYOUTRETURN and
We continue to test the ability of our prototype to send direct I/O data to data servers.
At the September 2006 NFSv4 Bake-a-thon, hosted by CITI, we continued to test the ability of CITI's Linux pNFS client to operate with multiple layouts, and the ability of CITI's Linux pNFS server to export pNFS capable underlying file systems.
We demonstrated the Linux pNFS client support for multiple layouts by copying files between multiple pNFS back ends.
The following pNFS implementations were tested.
- Clients: Linux, Solaris
- Servers: Network Appliance, Linux IBM GPFS, DESY dCache, Solaris, PVFS2
- Client: Linux
- Servers: Linux, Panasas
- Client: Linux
- Server: EMC
- Client: Linux
- Server: Linux
Our current Linux pNFS implementation uses a single whole file layout. We are extending the layout cache on the client and layout management on the server to support multiple layouts and small byte ranges.
In cooperation with EMC, we continue to develop a block layout driver module for the generic pNFS client.
We continue to measure I/O performance.
We joined the Ultralight project and are testing pNFS I/O using pNFS clients on 10 GbE against pNFS clusters on 1 GbE. The Linux pNFS client is included in the Ultralight kernel and distributed to Ultralight sites, providing opportunities for future long-haul WAN testing.
We are demonstrating pNFS file layout at SC06. The demonstration will include multiple 10G nic clients from the Sc06 demonstration floor accessing data across the ultralight network to an IBM GPFS cluster at CITI.
Task 2. Client Migration
Migration of client from one mount/metadata server to another to be demonstrated. This demonstration may be replicated at LANL depending on success of this work.
When a file system moves, the old server notifies clients with NFS4ERR_MOVED. Clients then reclaim state held on the old server by engaging in reboot recovery with the new server. For cluster file systems, server-to-server state transfer lets clients avoid the reclaim.
We redesigned state bookkeeping to ensure that state created on NFSv4 servers exporting the same cluster file system will not collide.
Server reboot recovery requires servers to save the clientid of active clients in stable storage. The present server implementation does this by writing directly to a filesystem via the vfs layer. A new server instance reads the state from stable storage, again directly via the vfs. We are rewriting this implementation to use a pipefs upcall/downcall interface instead of directly using the vfs layer, and are expanding the interface to support an upcall/downcall of all a clients in-memory state. The userland daemon can then support server-to-server state transfer to the cooresponding daemon on a new server. We have a prototype of the new upcall/down call interface, and have yet to prototype the server-to-server state transfer.
It remains to inform clients that state established with the old server remains valid on the new server. The IETF NFSv4 working group is considering solutions for the NFSv4.1 protocol, but NFSv4.0 clients will not have support for this feature. We will therefore need to provide Linux specific implementation support - perhaps a mount option or a /proc flag, or simply to try to use an old clientid against a new server on migration.
Task 3. Lock Analysis
Analysis of caching and lock coherency, demonstration of caching and lock performance with scaling, under various levels of conflict, using byte range locks (looking at lock splitting issues etc.).
The NFSv4 protocol supports three different lock-like operations: opens, byte-range locks, and delegations.
Unlike previous versions of NFS, NFSv4 has an on-the-wire OPEN operation.
The OPEN call includes the expected access mode, which may be read, write, or both. But it also includes a "deny" mode, which may be read, write, or both, or none. The server fails any open which whose access mode overlaps the deny mode of an existing open, or whose deny mode overlaps the access mode of an existing open.
Deny modes are not currently used by UNIX-like clients, our main focus, so we don't study this case.
However, all clients still do perform an OPEN each time an application opens a file, for several reasons: to ensure correct behavior in the presence of Windows clients, to request delegations, and to establish the state necessary to get posix byte-range locks, among other reasons.
All versions of NFS also tie data caching to open and close: data is flushed before close, and attributes revalidated before open, in such a way as to guarantee that the data seen after an open will always reflect writes any other client performed using file descriptors that were closed before the open.
POSIX byte-range locks are managed by applications using fcntl(). Each lock request has a byte-range and a type of read or write. Read locks conflict only with write locks, whereas write locks conflict with any other locks. Applications may perform read locks only on files which they have open for read, and write locks only on files which they have open for write.
Byte-range locks are normally advisory; that is, they do not conflict with IO operations. Such mandatory locking is supported by many unix-like operating systems, appears to be rarely used.
The NLM sideband protocol enables byte-range locks for versions of NFS earlier than NFSv4. NFSv4 incorporates byte-range locking into the main protocol. This makes it possible to support mandatory byte-range locking, but support for mandatory byte-range locking over NFSv4 is not supported by the linux implementation, and no support is planned at this time.
As with opens, byte-range locks also affect data caching: unlocks are not allowed to succeed until modified data in the locked range is written to the server, and locks must revalidate file data. Thus writes performed under a lock has been unlocked will be visible to any reader that locks the region after the unlock.
A server may optionally return a "delegation" with the response to any open call. Delegations may be of type read or write. Servers must guarantee that no client ever holds a read delegation for on a file that another client has open for write, or has a write delegation for. Similarly no client may hold a write delegation on a file that another client has open for read.
A server is never required to give out a delegation. Also, it may ask for the delegation back at any time, at which point the client is required to do what is necessary to establish on the server any opens or locks which it has performed locally before returning the delegation. Once returned, the client cannot regain the delegation without performing another open.
An NFS client is not normally synchronously notified of changes performed by another client, but as long as a client holds a delegation, the above rules guarantee that it will be. In theory it might be possible for applications to take advantage of this increased cache consistency. However, this is not useful in practice since a server is never required to give out a delegation. Also, a server can ask for a delegation back at any time.
Thus clients do not expose the existence of delegations to applications the way they do opens and locks. Instead, clients use delegations to provide increased performance: delegations allows clients to perform open and lock calls locally--in the case of a read delegation, read opens and read locks may be performed without contacting the server, and in the case of a write delegation, any opens and locks may be performed without contacting the server. This also relieves the client of the responsibility to flush dirty data and revalidate data caches.
When a server recalls a delegation, the client is required to perform opens, locks, and writes to the server as necessary to inform the server of any state that the client has established only locally. Conflicting opens will be delayed until this process is completed.
Lock performance test harness
In the performance measurements that follow, we used a single client and server. We also ran the experiments on the client hardware with the local file system for comparison.
- IBM/Lenovo Thinkpad T43
- 2GHz Pentium M CPU
- 512 MB RAM
- 1000bT NIC
- 5400 RPM Ultra-ATA 80GB HD
- running 2.6.17-CITI
- 1GHz Athlon 64 3000+ CPU
- 512 MB RAM
- 1000bT NIC
- 7200 RPM SATA-II 80GB HD
File lock measurements
Lacking examples of real-world lock-intensive workloads, we have performed a few microbenchmarks to measure such things as the cost of acquiring a single lock with and without a delegation.
To measure the performance of whole file locking, we use a benchmark that opens N files, then obtains a lock on each file. We measure the elapsed time of the loop that obtains the locks. We ran the microbenchmark on three configurations:
- Local (reiserFS) file system
- NFS without delegations
- NFS with delegations
To test with no delegations, we disabled file leasing, which disables delegations as a side effect.
For most cases, we ran the test 10 times and averaged the results. Variance was not negligible, so we show standard deviations.
With delegations enabled and 1,000 files, we average the result of two runs, not 10, because the server limited the number of delegations to something less than 3,000.
File lock measurement results
All times are in μsec.
Local file system
--- 1 lock, 10 runs --- 24 : mean time per lock to lock 1 file 27 : median 0.8 : std dev --- 10 locks, 10 runs --- 7.2 : mean time per lock to lock 10 files 6.9 : median 0.7 : std dev --- 100 locks, 10 runs --- 6.5 : mean time per lock to lock 100 files 6.4 : median 0.09: std dev --- 1000 locks, 10 runs --- 8.3 : mean time per lock to lock 1,000 files 6.6 : median 2.1 : std dev
NFS no read delegation
--- 1 lock, 10 runs --- 511 : mean time per lock to lock 1 file 303 : median 534 : std dev --- 10 locks, 10 runs --- 283 : mean time per lock to lock 10 files 267 : median 34.2 : std dev --- 100 locks, 10 runs --- 269 : mean time per lock to lock 100 files 266 : median 11.1 : std dev --- 1000 locks, 10 runs --- 305 : mean time per lock to lock 1,000 files 296 : median 34.4 : std dev
NFS w/ read delegation
--- 1 lock, 10 runs --- 2.8 : mean time per lock to lock 1 file 3.0 : median 6 : std dev --- 10 locks, 10 runs --- 9.1 : mean time per lock to lock 10 files 9.1 : median 0.0 : std dev --- 100 locks, 10 runs --- 7.6 : mean time per lock to lock 100 files 7.2 : median 1.1 : std dev --- 1,000 locks, 2 runs --- 8.0 : mean time per lock to lock 1,000 files 8.0 : median 0.049: std dev
File lock performance discussion
The pattern that emerges suggests that delegations improve the performance of whole-file locking. The cost per lock in resiserFS and in NFS with delegations is six to eight μsec per lock. Without delegations, the cost is greater by two orders of magnitude.
Further investigation is under way, examining the reduction in server load and in the number of RPCs when delegations are enabled.
Byte-range lock measurements
To measure the cost of byte-range locks, we focused on the cost of splitting and joining locks. (All locks discussed in this section are POSIX byte-range locks.)
For lock splitting, we created a 30 MB file, locked the entire file, then unlocked non-contiguous ranges. Each unlock operation splits the initial lock.
To measure the cost of lock joining, we ran a complementary test: we locked non-contiguous regions of the file, then locked the entire range. The ranges are non-contiguous to avoid coalescing locks as we proceed.
We measured performance for four segment sizes:
- 3 segments, each 100,000,000 bytes
- 30 segments, each 1,000,000 bytes
- 300 segments, each 100,000 bytes
- 3,000 segments, each 10,000 bytes
As before, we ran each test 10 times and averaged the results. Variance was negligible. Between runs, we unmounted and remounted the server to assure nothing was cached across runs.
We ran the join and split tests on the local file system, on NFS with no delegations, and again on NFS with delegations. After each test with delegations, we verified that the delegation was in place by opening the file for writing at the end of the run and noting the DELGETRETURN callback.
Byte-range lock split results
Local file system
0.000016 secs - lock whole file 0.000021 secs - unlock 3 10000000-byte regions (split) 0.000016 secs - lock whole file 0.000190 secs - unlock 30 1000000-byte regions (split) 0.000015 secs - lock whole file 0.002985 secs - unlock 300 100000-byte regions (split) 0.000015 secs - lock whole file 0.177999 secs - unlock 3000 10000-byte regions (split) 0.000016 secs - lock whole file 23.569079 secs - unlock 29971 1000-byte regions (split)
NFS, no delegations
0.000276 secs - lock whole file 0.000704 secs - unlock 3 10000000-byte regions (split) 0.000325 secs - lock whole file 0.007067 secs - unlock 30 1000000-byte regions (split) 0.000276 secs - lock whole file 0.073822 secs - unlock 300 100000-byte regions (split) 0.000271 secs - lock whole file 1.099715 secs - unlock 3000 10000-byte regions (split) 0.000289 secs - lock whole file 74.407294 secs - unlock 29971 1000-byte regions (split)
NFS with delegations
0.000016 secs - lock whole file 0.000026 secs - unlock 3 10000000-byte regions (split) 0.000016 secs - lock whole file 0.000248 secs - unlock 30 1000000-byte regions (split) 0.000017 secs - lock whole file 0.004350 secs - unlock 300 100000-byte regions (split) 0.000016 secs - lock whole file 0.225387 secs - unlock 3000 10000-byte regions (split) 0.000017 secs - lock whole file 22.961603 secs - unlock 29971 1000-byte regions (split)
Byte-range lock join results
Local file system
0.000028 secs - lock 3 10000000-byte regions 0.000010 secs - lock whole file (join) 0.000212 secs - lock 30 1000000-byte regions 0.000031 secs - lock whole file (join) 0.004222 secs - lock 300 100000-byte regions 0.000407 secs - lock whole file (join) 0.369237 secs - lock 3000 10000-byte regions 0.004469 secs - lock whole file (join) 43.966929 secs - lock 29971 1000-byte regions 0.030219 secs - lock whole file (join)
NFS, no delegations
0.000750 secs - lock 3 10000000-byte regions 0.000246 secs - lock whole file (join) 0.007616 secs - lock 30 1000000-byte regions 0.000307 secs - lock whole file (join) 0.081856 secs - lock 300 100000-byte regions 0.001215 secs - lock whole file (join) 1.548707 secs - lock 3000 10000-byte regions 0.011581 secs - lock whole file (join) 133.975178 secs - lock 29971 1000-byte regions 0.120294 secs - lock whole file (join)
NFS with delegations
0.000032 secs - lock 3 10000000-byte regions 0.000012 secs - lock whole file (join) 0.000284 secs - lock 30 1000000-byte regions 0.000046 secs - lock whole file (join) 0.006794 secs - lock 300 100000-byte regions 0.000558 secs - lock whole file (join) 0.347239 secs - lock 3000 10000-byte regions 0.002566 secs - lock whole file (join) 42.846999 secs - lock 29971 1000-byte regions 0.029043 secs - lock whole file (join)
Delegations work as advertised: they make the cost of lock splitting and joining approximately the same as the cost in the local file system.
Future testing should examine the cost of splits and joins over multiple clients, and the cost of random ordering of lock requests.
Although the lock join test requires non-contiguous ranges to avoid coalescing, the lock split test does not, and should be re-run with contiguous ranges.
Delegation recall with byte-range locks
Earlier, we saw the performance advantage for a client holding delegations when acquiring locks. Now we examine the performance penalty to the server when recalling a delegation in the face of numerous client locks.
To test the performance of delegation recall, the client opens a 30 MB file, acquires a number of byte-range locks, then idles. The server then opens the file for writing, which induces a delegation recall callback. This forces the client to release its locks locally and establish them on the server. The client then relinquishes the delegation. We measure the elapsed time for the server to process the open call.
We vary the number of locks, testing performance with 0, 1, 2, 3, 4, 5, 10, 25, 50, 100, 250, 500, 1,000, 2,500, 5,000, and 10,000 locks.
The following table shows the total open time measured on the server and the time normalized by the number of locks recalled.
Total times are measured in seconds. Normalized times are shown in msec.
We are investigating the unexpected nonlinear behavior in the cost per lock.
Task 4. Directory Delegations
Analysis of directory delegations – how well does it work and when, when does it totally not work.
Directory delegations promise to extend the usefulness of dentry caching in two ways. First, the client is no longer forced to revalidate the dentry cache after a timeout. Second, while positive caching can be treated as a hint, negative caching without cache invalidation violates open-to-close semantics. Directory delegations allow the client to cache negative results.
For example, if a client opens a file that does not exist, it issues an OPEN RPC that fails. But a subsequent open of the same file might succeed, if the file is created in the interim. Open-to-close semantics requires that the newly created file be seen by the client, so the earlier negative result can not be cached. Consequently, subsequent opens of the same non-existent file also require OPEN RPC calls being sent to the server. This example is played out repeatedly when the shell searches for executables in PATH or when the linker searches for shared libraries in LD_LIBRARY_PATH.
With directory delegations, the server callback mechanism can guarantee that no entries have been added or modified in a cached directory, which allows consistent negative caching and eliminates repeated checks for non-existent files.
We implemented directory delegations in the Linux NFSv4 client and server.
Our server implementation follows the file delegations architecture. We extended the lease API in the Linux VFS to support read-only leases on directories and NFS-specific lease-breaking semantics.
We implemented a /proc interface on the server to enable or disable directory delegation at run time. At startup, the client queries the server for directory delegation support.
The server has hooks for a policy layer to control the granting of directory delegations. (No policy is implemented yet.) When and whether to acquire delegations is also a client concern.
We are testing delegation grant and recall in a test rig with one or two clients. Testing consists mostly of comparing NFS operation-counts when directory delegations is enabled or disabled.
Tests range from simple UNIX utilities — ls, find, touch — to hosting a CVS repository or compiling with shared libraries and header files on NFS servers. Tests will become more specific.
We have extended PyNFS to support directory delegations. So far, the support is basic and the tests are trivial. Tests will become more specific.
We are designing mechanisms that allow simulation experiments to compare delegation policies on NFSv4 network traces.
A comprehensive report on CITI experience with directory delegationsis under preparation. We will send that report when it is complete.
Task 5. NFS Server Load
How do you specify/measure NFS Server load?
To frame the task, consider identical symmetric servers with a cluster file system back end and a task running on one of them. Can we compare the load on the servers to determine whether there would be a benefit to migrating a client from one to the other?
Answering this question requires that we define a model of load based on measurable quanta.
Given a model, the next step is to write a tool that collects the factors that influence load and to measure how well the model accurately predicts performance.
If an application is running at less than peak performance, the load model should tell us whether the bottleneck is in the server, the client, or elsewhere.
If the bottleneck is in the server, one option for improving application performance is replacing server components with faster ones. Another option is to add servers. A third option is to migrate the application to a lightly-loaded server.
Factors that influence server load
The hardware characteristics that influence server capacity include disk bandwidth, CPU speed, interrupt rate, and memory availability.
The rate at which a single file in a server file system can be read from or written to depends on many factors, including characteristics of the disk hardware (rotation speed, access latency, etc.), the disk controller, the bus, the layout of the file on disk, the size of the transfer, and the degree of caching. The overall bandwidth of a file system also depends on the degree of striping and distribution of requests across disks.
The iostat command can reveal a bottleneck due to server disks if seek or transfer rates approach maximum values. For a given server configuration, these values can be measured directly. It might be possible to predict these values for a given hardware ensemble.
Server threads compete with one another and with the operating system for access to the CPU. Excess offered load can exhaust the availability of server threads.
* how would we know if this were to happen? * would it suffice to simply allocate more threads? * or are there pathological cases to consider?
Overall CPU utilization can be measured, also with iostat, but there may be other factors influencing the allocation of CPU to server threads. For example, excessive pressure on the memory or interrupt subsystem can force the operating system to intervene.
We're conflating thread usage and CPU usage. The two aren't the same, since nfsd threads can block on IO. So a sudden storm of "commit" requests, for example, (which require IO to be sync'd to disk) could tie up every nfsd thread while leaving the CPU idle.
So to do this right we'd need a separate section for threads, alongside CPU, disks, etc. I'd rather not do that, because it treats this behavior as an intrinsic feature of the server rather than as a bug to be fixed (which is what it probably is). So I'd be inclined just to omit any mention of threads here....
Interrupt rates can be measured with the vmstat command.
For a given hardware configuration, a threshold can be measured experimentally.
The memory subsystem is complex and varies among operating systems. Applications compete with one another for virtual memory. Often, they also compete with the file system, which uses the virtual memory subsystem for its in-memory cache.
Often, excess demand for memory is reflected by early eviction of pages in virtual memory. The vmstat command shows the pageout rate, which does not measure early eviction, but does reflect overall memory pressure.
Network utilization is the ratio of delivered bandwidth to maximum available bandwidth. Maximum available bandwidth is a property of network hardware. Delivered bandwidth can be measured with the netstat command.
Full-duplex network technologies can deliver maximal bandwidth in both directions, while half-duplex network technologies are limited to delivering the sum of the two directions.
* i believe that is a true statement ...
The overall performance of a server can be tested by measuring NFS performance directly with microbenchmarks. Candidate microbenchmarks include NULL RPC, and small READ RPC, large READ RPC, small WRITE RPC, and large WRITE RPC. The latency of a single operation or the maximum rate at which repeated operations can be performed are of interest.
Macrobenchmarks might also be useful, such as an application that performs specific tasks. Many popular benchmarks unpack and build an application like SSH or the Linux kernel.
The usefulness of a measured value can be tested by comparing microbenchmark performance as the resource is consumed. For example:
- We might vary the size of a raid array to reduce the overall disk bandwidth.
- We can run a process that consumes varying amounts of CPU time.
- We can vary the number of threads in the server pool.
- We can vary the number of connections that can be served by a single thread.
- We can run a pair of processes that use varying amounts of network bandwidth.
Each measured value can be expressed as a ratio between 0 (idle) and 1 (at capacity). For each value, there is a program that consumes the corresponding resource. We can then compare the measured server performance as the amount of the resource is varied.
It is also useful to sample the instantaneous values, and to track them over time with a damping function that shows the averages over the last second, minute, five minutes, etc.
Some special situations might be used to measure server performance:
- Reboot recovery: measure elapsed time when a number of clients are simultaneously recovering.
- Mount storms: measure the elapsed time when a large number of clients simultaneosly attempt to mount a server. This might arise when a cluster job triggers an automount from all clients at once.
Possible benchmark sources
Some macrobenchmarks are already in common use.
- Postmark is a popular macrobenchmark that simulates the operation of a mail server with a mix of reads, writes, creates, and unlinks. Postmark does not exercise locks.
- Filebench also lacks locks. We haven't figured out exactly what the various loads do and are trying to find the relevant developer community.
- IOzone can use locks.
- LLNL and other labs have some metadata stress tests, e.g., IOR and mdtest. IOR places enormous stress on a server, and scales to thousands of clients.
- ASCI Purple Benchmark Codes
- Resources at Dartmouth
- Bull.net uses Bonnie++, FStress, and dbench in its NFS load testing. dbench simulates the file system activity created by a samba server running the proprietary SMB benchmark netbench.