Thanks for your great work! +1 for merging phase 1 codes. My product clusters have been running phase 1 codes for several months, it looks good.
Hope to push this feature forward. | | 张浩博 | | hfutzhan...@163.com | ---- Replied Message ---- | From | haiyang hu<haiyang87...@gmail.com> | | Date | 12/31/2024 23:08 | | To | Ayush Saxena<ayush...@gmail.com> | | Cc | Hui Fei<feihui.u...@gmail.com> , ZanderXu<zande...@apache.org> , Hdfs-dev<hdfs-dev@hadoop.apache.org> , <priv...@hadoop.apache.org> , Xiaoqiao He<hexiaoq...@apache.org> , slfan1989<slfan1...@apache.org> , <xuzq_zan...@163.com> | | Subject | Re: Discussion about NameNode Fine-grained locking | Thanks for your hard work and push it forward. It looks good, +1 for merging phase 1 codes, hope we can work together to promote this major HDFS optimization, so that more companies can benefit from it. Thanks everyone~ Ayush Saxena <ayush...@gmail.com> 于2024年12月31日周二 20:33写道： +1, Thanx folks for your efforts on this! I didn't have time to review everything thoroughly, but my initial pass suggests it looks good or atleast is safe to merge. If I find some spare time, I'll test it further and submit a ticket or so if I encounter any issues. Good Luck!!! -Ayush On Tue, 31 Dec 2024 at 16:39, Hui Fei <feihui.u...@gmail.com> wrote: Thanks Zander for bringing this discussion again and trying your best to push it forward. It's really a long time since last discussion. It’s indeed time, +1 for merging phase 1 codes based on the following points - The phase 1 feature has been running at scale within companies for a long time - The long-term plan is clear, and also addressed some questions raised by the community - The testing result of future features on memory and performance ZanderXu <zande...@apache.org> 于2024年12月31日周二 15:36写道： Hi, everyone: Time to Merge FGL Phase I The PR for FGL Phase I is ready for merging! Please take a moment to review and cast your vote: https://github.com/apache/hadoop/pull/6762. The FGL Phase I has been running successfully in production for over six months at Shopee and BOSS Zhipin, with no reported performance or stability issues. It’s now the right time to merge it into the trunk branch, allowing us to move forward with Phase II. The global lock remains the default lock mode, but users can enable FGL by configuring dfs.namenode.lock.model.provider.class=org.apache.hadoop.hdfs.server.namenode.fgl.FineGrainedFSNamesystemLock. If there are no objections within 7 days, I will propose an official vote. Performance and Memory Usage of Phase I Conclusion： Fine-grained locks do not lead to significant performance improvements. Fine-grained locks do not result in additional memory consumption Reasons: BM operations heavily depend on FS operations: IBR and BR still acquire the global lock (FSLock and BMLock). FS operations depend on BM operations: Common operations (create, addBlock, getBlockLocations) also acquire the global lock (FSLock and BMLock). Phase II will bring significant performance improvements by decoupling FS and BM dependencies and replacing the global FSLock with a fine-grained IIPLock. Addressing Common Questions Thank you all for raising meaningful questions! I have rewritten the design document to improve clarity. https://docs.google.com/document/d/1DXkiVxef9wCmICjpZyIQO-yxsgwc4wnf2lTKQ3UXe30/edit?usp=sharing Below is a summary of frequently asked questions and answers: Summary of Questions: Question 1: How is the performance of LockPoolManager? Performance Report: Time to acquire a cached lock: 194 ns Time to acquire a non-cached lock: 1044 ns Time to release an in-use lock: 88 ns Time to release an unused lock: 112 ns Overall Performance: QPS: Over 10 million Time to acquire the IIP lock for a path with depth 10: Fully uncached: 10440 ns + 1120 ns (≈ 11 μs) Fully cached: 1940 ns + 1120 ns (≈ 3 μs) In global lock scenarios, lock wait times are typically in the millisecond range. Therefor, the cost of acquiring and releasing fine-grained locks can be ignored. Question 2: How much memory does the FGL consume? Memory Consumption: A single LockResource contains a read-write lock and a counter, totaling approximately 200 bytes: LockResource: 24 bytes ReentrantReadWriteLock: 150 bytes AtomicInteger: 16 bytes Memory Usage Estimates: 10-level directory depth, 100 handlers 1000 lock resources, approximately 200 KB 10-level directory depth, 1000 handlers 10000 lock resources, approximately 2 MB 1, 000,000 lock resources, approximately 200 MB Conclusion: Memory consumption is negligible. Question 3: What happens if no lock is available in the LockPoolManager? If there are not any available LockResources, two solutions are available: Return a RetryException, prompting the client to retry later. Temporarily increase the lock entity limit, allocate more locks to meet client requests, and use an asynchronous thread to recycle locks periodically. We can provide multiple LockPoolManager implementations for users to choose from based on production environments. Question 4: Regarding the IIPLock lock depth issue, can we consider holding only the first 3 or 4 levels of directory locks? This approach is not recommended for the following reasons: Cannot maximize concurrency. Limited savings in lock acquisition/release time and memory usage, yielding insignificant benefits. Question 5: How should attributes like StoragePolicy, ErasureCoding, and ACL, which can be set on parent or ancestor directory nodes, be handled? ErasureCoding and ACL: When changing node attributes, hold the corresponding INode’s write lock. When using ancestor node attributes, hold the corresponding INode’s read lock. StoragePolicy: More complex due to its impact on both directory tree operations and Block operations. To improve performance, commonly used block-related operations (such as BR/IBR) should not acquire IIPLock Detailed design documentation: https://docs.google.com/document/d/1DXkiVxef9wCmICjpZyIQO-yxsgwc4wnf2lTKQ3UXe30/edit?tab=t.0#heading=h.96lztsl4mwfk Question 6: How should FGL be implemented for the SNAPSHOT feature? Since the Rename operation on the SNAPSHOT directory is supported, holding only the write lock of the SNAPSHOT root directory cannot cover the rename situation, so the thread safety of SNAPSHOT-related operations cannot be guaranteed It is recommended to use global FS lock to ensure thread safety. Detailed design documentation: https://docs.google.com/document/d/1DXkiVxef9wCmICjpZyIQO-yxsgwc4wnf2lTKQ3UXe30/edit?tab=t.0#heading=h.sm36p6bfcpec Question 7: How should FGL be implemented for the Symlinks feature? The Target path of Symlinks is a string, and the client performs a second forward access to the Target path. So the fine-grained lock project requires no special handling For the createSymlink RPC, the FGL needs to acquire the IIPLocks for both target and link paths. Question 8: How should FGL be implemented for the reserved feature? The Reserved feature has two usage modes: /.reserved/iNodes/${inode id} /.reserved/raw/${path} INodeId Mode: During the resolvePath phase, obtain the real IIPLock lock via INodeId. Path Mode: During the resolvePath phase, obtain the real IIPLock lock via path. Detailed design documentation: https://docs.google.com/document/d/1DXkiVxef9wCmICjpZyIQO-yxsgwc4wnf2lTKQ3UXe30/edit?tab=t.0#heading=h.h6rcpzkbpanf Question 9: Why is INodeFileLock used as the FGL for BlockInfo? INodeFile and Block have mutual dependencies: INodeFile depends on Block for state and size. Block depends on INodeFile for state and storage policy. Therefore, using INodeFileLock as the fine-grained lock for BlockInfo is a reasonable choice. Detailed design documentation: https://docs.google.com/document/d/1DXkiVxef9wCmICjpZyIQO-yxsgwc4wnf2lTKQ3UXe30/edit?tab=t.0#heading=h.zesd6omuu3kr Seeking Community Feedback Your questions and concerns are always welcome. We can discuss them in detail on the Slack Channel: https://app.slack.com/client/T4S1WH2J3/C06UDTBQ2SH Let’s work together to advance the Fine-Grained Lock project. I believe this initiative will deliver significant performance improvements to the HDFS community and help reinvigorate its activity. Wishing everyone a Happy New Year 2025! On Wed, 5 Jun 2024 at 16:17, ZanderXu <zande...@apache.org> wrote: I plan to hold a meeting on 2024-06-06 from 3:00 PM - 4:00 PM to share the FGL's motivations and some concerns in detail in Chinese. The doc is : NameNode Fine-Grained Locking Based On Directory Tree (II) The meeting URL is: https://sea.zoom.us/j/94168001269 You are welcome to this meeting. On Mon, 6 May 2024 at 23:57, Hui Fei <feihui.u...@gmail.com> wrote: BTW, there is a Slack channel hdfs-fgl for this feature. can join it and discuss more details. Is it necessary to hold a meeting to discuss this? So that we can push it forward quickly. Agreed with ZanderXu, it seems inefficient to discuss details via email list. Hui Fei <feihui.u...@gmail.com> 于2024年5月6日周一 23:50写道： Thanks all Seems all concerns are related to the stage 2. We can address these and make it more clear before we start it. From development experience, I think it is reasonable to split the big feature into several stages. And stage 1 is also independent and it also can be as a minor feature that uses fs and bm locks instead of the global lock. ZanderXu <zande...@apache.org> 于2024年4月29日周一 15:17写道： Thanks @Ayush Saxena <ayush...@gmail.com> and @Xiaoqiao He <hexiaoq...@apache.org> for your nice questions. Let me summarize your concerns and corresponding solutions: *1. Questions about the Snapshot feature* It's difficult to apply the FGL to Snapshot feature, but we can just using the global FS write lock to make it thread safe. So if we can identity if a path contains the snapshot feature, we can just using the global FS write lock to protect it. You can refer to HDFS-17479 <https://issues.apache.org/jira/browse/HDFS-17479> to get how to identify it. Regarding performance of the operations related to the snapshot features, we can discuss it in two categories: Read operations involves snapshots: The FGL branch uses the global write lock to protect them, the GLOBAL branch uses the global read lock to protect them. It's hard to conclude which version has better performance, it depends on the global lock competition. Write operations involves snapshots: Both FGL and GLOBAL branch use the global write lock to protect them. It's hard to conclude which version has better performance, it depends on the global lock competition too. So I think if namenode load is low, the GLOBAL branch will have a better performance than FGL; If namenode load is high, the FGL branch may have a better performance than the GLOBAL, which also depends on the ratio of read and write operations on the SNAPSHOT feature. We can do somethings to let end-user to choose a branch with a better branch according to their business: First, we need to make the lock mode can be selectable, so that end-user can choose to use FGL of GLOBAL. Second, using the global write lock to make operations related to snapshot thread safe as I described in HDFS-17479. *2. Questions about the Symlinks feature* If Symlink is related to snapshot, we can refer to the solution of the snapshot; If Symlink is not related to snapshot, I think it's easy to meet the FGL. Only createSymlink involves two paths, FGL just need to lock them in the order to make this operation thread. For other operations, it is the same as other normal iNode, right? If I missed difficult points, please let me know. *3. Questions about Memory Usage of iNode locks* I think there are too many solutions to limit the memory usage of these iNode locks, such as: Using a limit capacity lock pool to ensure the maximum memory usage, Just holding iNode locks for fixed depth of directories, etc. We can just abstract this LockManager first and then support its implementation with different ideas, so that we can limit the maximum memory usage of these iNode locks. FGL can acquire or lease iNode locks through LockManager. *4. Questions about Performance of acquiring and releasing iNode locks* We can add some benchmark for LockManager, to test the performance or acquire and release unblocked locks. *5. Questions about StoragePolicy, ECPolicy, ACL, Quota, etc.* These policies may be sot on an ancestor node and used by some children files. The set operation for these policies will be protected by the directory tree, since there are all file-related operations. In addition to Quota and StoragePolicy, the use of other policies will also be protected by directory tree, such as ECPolicy and ACL. Quota is a little special since its update operations may not be protected by the directory tree, we can assign a locks to each QuotaFeature and use these locks to make updating operations thread safe. you can refer to HDFS-17473 <https://issues.apache.org/jira/browse/HDFS-17473> to get some detailed information. StoragePolicy is a little special since it is used not only by file-related operations but also block-related operations. ProcessExtraRedundancyBlock uses storage policy to choose redundancy replicas and BlockReconstructionWork uses storage policy to choose target DNs. In order to maximize the performance improvement, BR and IBR should only involve the iNodeFile to which the current processing block belongs. These redundancy blocks can be processed by the Redundancy monitor while holding the directory tree locks. You can refer to HDFS-17505 <https://issues.apache.org/jira/browse/HDFS-17505> to get more detailed informations. *6. Performance of the phase 1* HDFS-17506 <https://issues.apache.org/jira/browse/HDFS-17506> is used to do some performance testing for phase 1, and I will complete it later. Discuss solution through mails is not efficient, you can create one sub-tasks under HDFS-17366 <https://issues.apache.org/jira/browse/HDFS-17366> to describe your concerns and I will try to give some answers. Thanks @Ayush Saxena <ayush...@gmail.com> and @Xiaoqiao He <hexiaoq...@apache.org> again. On Mon, 29 Apr 2024 at 02:00, Ayush Saxena <ayush...@gmail.com> wrote: Thanx Everyone for chasing this, Great to see some momentum around FGL, that should be a great improvement. I have some two broad categories: ** About the process:* I think in the above mails, there are mentions that phase one is complete in a feature branch & we are gonna merge that to trunk. If I am catching it right, then you can't hit the merge button like that. To merge a feature branch. You need to call for a Vote specific to that branch & it requires 3 binding votes to merge, unlike any other code change which requires 1. It is there in our Bylaws. So, do follow the process. ** About the feature itself:* (A very quick look at the doc and the Jira, so please take it with a grain of salt) * The Google Drive link that you folks shared as part of the first mail. I don't have access to that. So, please open up the permissions for that doc or share the new link * Chasing the design doc present on the Jira * I think we only have Phase-1 ready, so can you share some metrics just for that? Perf improvements just with splitting the FS & BM Locks * The memory implications of Phase-1? I don't think there should be any major impact on the memory in case of just phase-1 * Regarding the snapshot stuff, you mentioned taking lock on the root itself? Does just taking lock on the snapshot root rather than the FS root works? * Secondly about the usage of Snapshot or Symlinks, I don't think we should operate under the assumptions that they aren't widely used or not, we might just not know folks who don't use it widely or they are just users not the ones contributing. We can just accept for now, that in those cases it isn't optimised and we just lock the entire FS space, which it does even today, so no regressions there. * Regarding memory usage: Do you have some numbers on how much the memory footprint increases? * Under the Lock Pool: I think you are assuming there would be very few inodes where lock would be required at any given time, so there won't be too much heap consumption? I think you are compromising on the Horizontal Scalability here. I doubt if your assumption doesn't hold true, under heavy read load by concurrent clients accessing different inodes, the Namenode will start giving memory troubles, that would do more harm than good. Anyway Namenode heap is way bigger problem than anything, so we should be very careful increasing load over there. * For the Locks on the inodes: Do you plan to have locs for each inode? Can we somehow limit that to the depth of the tree? Like currently we take lock on the root, have a config which makes us take lock at Level-2 or 3 (configurable), that might fetch some perf benefits and can be used to control the memory usage as well? * What is the cost of creating these inode locks? If the lock isn't already cached it would incur some cost? Do you have some numbers around that? Say I disable caching altogether & then let a test load run, what does the perf numbers look like in that case * I think we need to limit the size of INodeLockPool, we can't let it grow infinitely in case of heavy loads and we need to have some auto throttling mechanism for it * I didn't catch your Storage Policy problem. If I decode it right, the problem is like the policy could be set on an ancestor node & the children abide by that & this is the problem, if that is the case then isn't that the case with ErasureCoding policies or even ACLs or so? Can you elaborate a bit on that. Anyway, regarding the Phase-1. If you share (the perf numbers with proper details + Impact on memory if any) for just phase 1 & if they are good, then if you call for a branch merge vote for Phase-1 FGL, you have my vote, however you'll need to sway the rest of the folks on your own :-) Good Luck, Nice Work Guys!!! -Ayush On Sun, 28 Apr 2024 at 18:32, Xiaoqiao He <hexiaoq...@apache.org> wrote: Thanks ZanderXu and Hui Fei for your work on this feature. It will be a very helpful improvement for the HDFS module in the next journal. 1. If we need any more review bandwidth, I would like to be involved to help review if possible. 2. From the design document there are still missing some detailed descriptions such as snapshot, symbolic link and reserved etc as mentioned above. I think it will be helpful for newbies who want to be involved if all corner cases are considered and described. 3. From slack, we plan to check into the trunk at this phase. I am not sure If it is the proper time, following the dev plan there are two steps left to finish this feature from the design document, right? If that, I think we should postpone checking in when all plans are ready. Considering that there are many unfinished tries for this feature in history, I think postpone checking will be the safe way, another way it will involve more rebase cost if you keep separate dev branch, however I think It is not one difficult thing for you. Good luck and look forward to making that happen soon! Best Regards, - He Xiaoqiao On Fri, Apr 26, 2024 at 3:50 PM Hui Fei <feihui.u...@gmail.com> wrote: Thanks for interest and advice on this. Just would like to share some info here ZanderXu leads this feature and he has spent a lot of time on it. He is the main developer in stage 1. Yuanboliu and Kokonguyen191 also took some tasks. Other developers (slfan1989 haiyang1987 huangzhaobo99 RocMarshal kokonguyen191) helped review PRs. (Forgive me if I missed someone) Actually haiyang1987, Yuanboliu and Kokonguyen191 are also very familiar with this feature. We discussed many details offline. Welcome to more people interested in joining the development and review of the stage 2 and 3. Zengqiang XU <xuzengqiang5...@gmail.com> 于2024年4月26日周五 14:56写道： Thanks Shilun for your response: 1. This is a big and very useful feature, so it really needs more developers to get on board. 2. This fine grained lock has been implemented based on internal branches and has gained benefits by many companies, such as: Meituan, Kuaishou, Bytedance, etc. But it has not been contributed to the community due to various reasons, such as there is a big difference between the version of the internal branch and the community trunk branch, the internal branch may ignore some functions to make FGL clear, and the contribution needs a lot of work and will take many times. It means that this solution has already been practiced in their prod environment. We have also practiced it in our prod environment and gained benefits, and we are also willing to spend a lot of time contributing to the community. 3. Regarding the benchmark testing, we don't need to pay more attention to whether the performance is improved by 5 times, 10 times or 20 times, because there are too many factors that affect it. 4. As I described above, this solution is already being practiced by many companies. Right now, we just need to think about how to implement it with high quality and more comprehensively. 5. I firmly believe that all problems can be solved as long as the overall solution is right. 6. I can spend a lot of time leading the promotion of this entire feature and I hope more people can join us in promoting it. 7. You are always welcome to raise your concerns. Thanks Shilun again, I hope you can help review designs and PRs. Thanks On Fri, 26 Apr 2024 at 08:00, slfan1989 <slfan1...@apache.org> wrote: Thank you for your hard work! This is a very meaningful improvement, and from the design document, we can see a significant increase in HDFS read/write throughput. I am happy to see the progress made on HDFS-17384. However, I still have some concerns, which roughly involve the following aspects: 1. While ZanderXu and Hui Fei have deep expertise in HDFS and are familiar with related development details, we still need more community member to review the code to ensure that the relevant upgrades meet expectations. 2. We need more details on benchmarks to ensure that test results can be reproduced and to allow more community member to participate in the testing process. Looking forward to everything going smoothly in the future. Best Regards, - Shilun Fan. On Wed, Apr 24, 2024 at 3:51 PM Xiaoqiao He < hexiaoq...@apache.org> wrote: cc private@h.a.o. On Wed, Apr 24, 2024 at 3:35 PM ZanderXu < zande...@apache.org> wrote: Here are some summaries about the first phase: 1. There are no big changes in this phase 2. This phase just uses FS lock and BM lock to replace the original global lock 3. It's useful to improve the performance, since some operations just need to hold FS lock or BM lock instead of the global lock 4. This feature is turned off by default, you can enable it by setting dfs.namenode.lock.model.provider.class to org.apache.hadoop.hdfs.server.namenode.fgl.FineGrainedFSNamesystemLock 5. This phase is very import for the ongoing development of the entire FGL Here I would like to express my special thanks to @kokonguyen191 and @yuanboliu for their contributions. And you are also welcome to join us and complete it together. On Wed, 24 Apr 2024 at 14:54, ZanderXu < zande...@apache.org> wrote: Hi everyone All subtasks of the first phase of the FGL have been completed and I plan to merge them into the trunk and start the second phase based on the trunk. Here is the PR that used to merge the first phases into trunk: https://github.com/apache/hadoop/pull/6762 Here is the ticket: https://issues.apache.org/jira/browse/HDFS-17384 I hope you can help to review this PR when you are available and give some ideas. HDFS-17385 < https://issues.apache.org/jira/browse/HDFS-17385> is used for the second phase and I have created some subtasks to describe solutions for some problems, such as: snapshot, getListing, quota. You are welcome to join us to complete it together. ---------- Forwarded message --------- From: Zengqiang XU <zande...@apache.org> Date: Fri, 2 Feb 2024 at 11:07 Subject: Discussion about NameNode Fine-grained locking To: <hdfs-dev@hadoop.apache.org> Cc: Zengqiang XU <xuzengqiang5...@gmail.com> Hi everyone I have started a discussion about NameNode Fine-grained Locking to improve performance of write operations in NameNode. I started this discussion again for serval main reasons: 1. We have implemented it and gained nearly 7x performance improvement in our prod environment 2. Many other companies made similar improvements based on their internal branch. 3. This topic has been discussed for a long time, but still without any results. I hope we can push this important improvement in the community so that all end-users can enjoy this significant improvement. I'd really appreciate you can join in and work with me to push this feature forward. Thanks very much. Ticket: HDFS-17366 < https://issues.apache.org/jira/browse/HDFS-17366> Design: NameNode Fine-grained locking based on directory tree < https://docs.google.com/document/d/1X499gHxT0WSU1fj8uo4RuF3GqKxWkWXznXx4tspTBLY/edit?usp=sharing --------------------------------------------------------------------- To unsubscribe, e-mail: private-unsubscr...@hadoop.apache.org For additional commands, e-mail: private-h...@hadoop.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org
