Message ID | 1536253938-192391-1-git-send-email-honnappa.nagarahalli@arm.com |
---|---|
Headers | show |
Series | Address reader-writer concurrency in rte_hash | expand |
I have added the memory ordering ladder diagrams to the DPDK summit slides to help understand the changes. The slides are available at: https://dpdkuserspace2018.sched.com/event/G44w/lock-free-read-write-concurrency-in-rtehash. Please look at the backup slides.
Thank you,
Honnappa
-----Original Message-----
From: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Sent: Thursday, September 6, 2018 12:12 PM
To: bruce.richardson@intel.com; pablo.de.lara.guarch@intel.com
Cc: dev@dpdk.org; honnappa.nagarahalli; Gavin Hu (Arm Technology China) <Gavin.Hu@arm.com>; Steve Capper <Steve.Capper@arm.com>; Ola Liljedahl <Ola.Liljedahl@arm.com>; nd <nd@arm.com>; Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>
Subject: [PATCH 0/4] Address reader-writer concurrency in rte_hash
Currently, reader-writer concurrency problems in rte_hash are
addressed using reader-writer locks. Use of reader-writer locks
results in following issues:
1) In many of the use cases for the hash table, writer threads
are running on control plane. If the writer is preempted while
holding the lock, it will block the readers for an extended period
resulting in packet drops. This problem seems to apply for platforms
with transactional memory support as well because of the algorithm
used for rte_rwlock_write_lock_tm:
static inline void
rte_rwlock_write_lock_tm(rte_rwlock_t *rwl)
{
if (likely(rte_try_tm(&rwl->cnt)))
return;
rte_rwlock_write_lock(rwl);
}
i.e. there is a posibility of using rte_rwlock_write_lock in
failure cases.
2) Reader-writer lock based solution does not address the following
issue.
rte_hash_lookup_xxx APIs return the index of the element in
the key store. Application(reader) can use that index to reference
other data structures in its scope. Because of this, the
index should not be freed till the application completes
using the index.
3) Since writer blocks all the readers, the hash lookup
rate comes down significantly when there is activity on the writer.
This happens even for unrelated entries. Performance numbers
given below clearly indicate this.
Lock-free solution is required to solve these problems. This patch
series adds the lock-free capabilities in the following steps:
1) Correct the alignment for the key store entry to prep for
memory ordering.
2) Add memory ordering to prevent race conditions when a new key
is added to the table.
3) Reader-writer concurrency issue, caused by moving the keys
to their alternate locations during key insert, is solved
by introducing an atomic global counter indicating a change
in table.
4) This solution also has to solve the issue of readers using
key store element even after the key is deleted from
control plane.
To solve this issue, the hash_del_key_xxx APIs do not free
the key store element. The key store element has to be freed
using the newly introduced rte_hash_free_key_with_position API.
It needs to be called once all the readers have stopped using
the key store element. How this is determined is outside
the scope of this patch (RCU is one such mechanism that the
application can use).
4) Finally, a lock free reader-writer concurrency flag is added
to enable this feature at run time.
Performance numbers:
Scenario: Equal number of writer/reader threads for concurrent
writers and readers. For readers only test, the
entries are added upfront.
Current code:
Cores Lookup Lookup
with add
2 474 246
4 935 579
6 1387 1048
8 1766 1480
10 2119 1951
12 2546 2441
With this patch:
Cores Lookup Lookup
with add
2 291 211
4 297 196
6 304 198
8 309 202
10 315 205
12 319 209
Honnappa Nagarahalli (4):
hash: correct key store element alignment
hash: add memory ordering to avoid race conditions
hash: fix rw concurrency while moving keys
hash: enable lock-free reader-writer concurrency
lib/librte_hash/rte_cuckoo_hash.c | 445 +++++++++++++++++++++++++----------
lib/librte_hash/rte_cuckoo_hash.h | 6 +-
lib/librte_hash/rte_hash.h | 63 ++++-
lib/librte_hash/rte_hash_version.map | 7 +
4 files changed, 393 insertions(+), 128 deletions(-)
--
2.7.4
Hi Bruce/Pablo, I need to get this into 18.11, appreciate any review/feedback soon. Thank you, Honnappa > -----Original Message----- > From: Honnappa Nagarahalli > Sent: Friday, September 14, 2018 4:19 PM > To: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>; > bruce.richardson@intel.com; pablo.de.lara.guarch@intel.com > Cc: dev@dpdk.org; honnappa.nagarahalli; Gavin Hu (Arm Technology China) > <Gavin.Hu@arm.com>; Steve Capper <Steve.Capper@arm.com>; Ola Liljedahl > <Ola.Liljedahl@arm.com>; nd <nd@arm.com>; yipeng1.wang@intel.com; > Michel Machado <michel@digirati.com.br>; sameh.gobriel@intel.com > Subject: RE: [PATCH 0/4] Address reader-writer concurrency in rte_hash > > I have added the memory ordering ladder diagrams to the DPDK summit slides > to help understand the changes. The slides are available at: > https://dpdkuserspace2018.sched.com/event/G44w/lock-free-read-write- > concurrency-in-rtehash. Please look at the backup slides. > > Thank you, > Honnappa > > -----Original Message----- > From: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com> > Sent: Thursday, September 6, 2018 12:12 PM > To: bruce.richardson@intel.com; pablo.de.lara.guarch@intel.com > Cc: dev@dpdk.org; honnappa.nagarahalli; Gavin Hu (Arm Technology China) > <Gavin.Hu@arm.com>; Steve Capper <Steve.Capper@arm.com>; Ola Liljedahl > <Ola.Liljedahl@arm.com>; nd <nd@arm.com>; Honnappa Nagarahalli > <Honnappa.Nagarahalli@arm.com> > Subject: [PATCH 0/4] Address reader-writer concurrency in rte_hash > > Currently, reader-writer concurrency problems in rte_hash are > addressed using reader-writer locks. Use of reader-writer locks > results in following issues: > > 1) In many of the use cases for the hash table, writer threads > are running on control plane. If the writer is preempted while > holding the lock, it will block the readers for an extended period > resulting in packet drops. This problem seems to apply for platforms > with transactional memory support as well because of the algorithm > used for rte_rwlock_write_lock_tm: > > static inline void > rte_rwlock_write_lock_tm(rte_rwlock_t *rwl) > { > if (likely(rte_try_tm(&rwl->cnt))) > return; > rte_rwlock_write_lock(rwl); > } > > i.e. there is a posibility of using rte_rwlock_write_lock in > failure cases. > 2) Reader-writer lock based solution does not address the following > issue. > rte_hash_lookup_xxx APIs return the index of the element in > the key store. Application(reader) can use that index to reference > other data structures in its scope. Because of this, the > index should not be freed till the application completes > using the index. > 3) Since writer blocks all the readers, the hash lookup > rate comes down significantly when there is activity on the writer. > This happens even for unrelated entries. Performance numbers > given below clearly indicate this. > > Lock-free solution is required to solve these problems. This patch > series adds the lock-free capabilities in the following steps: > > 1) Correct the alignment for the key store entry to prep for > memory ordering. > 2) Add memory ordering to prevent race conditions when a new key > is added to the table. > > 3) Reader-writer concurrency issue, caused by moving the keys > to their alternate locations during key insert, is solved > by introducing an atomic global counter indicating a change > in table. > > 4) This solution also has to solve the issue of readers using > key store element even after the key is deleted from > control plane. > To solve this issue, the hash_del_key_xxx APIs do not free > the key store element. The key store element has to be freed > using the newly introduced rte_hash_free_key_with_position API. > It needs to be called once all the readers have stopped using > the key store element. How this is determined is outside > the scope of this patch (RCU is one such mechanism that the > application can use). > > 4) Finally, a lock free reader-writer concurrency flag is added > to enable this feature at run time. > > Performance numbers: > Scenario: Equal number of writer/reader threads for concurrent > writers and readers. For readers only test, the > entries are added upfront. > > Current code: > Cores Lookup Lookup > with add > 2 474 246 > 4 935 579 > 6 1387 1048 > 8 1766 1480 > 10 2119 1951 > 12 2546 2441 > > With this patch: > Cores Lookup Lookup > with add > 2 291 211 > 4 297 196 > 6 304 198 > 8 309 202 > 10 315 205 > 12 319 209 > > Honnappa Nagarahalli (4): > hash: correct key store element alignment > hash: add memory ordering to avoid race conditions > hash: fix rw concurrency while moving keys > hash: enable lock-free reader-writer concurrency > > lib/librte_hash/rte_cuckoo_hash.c | 445 +++++++++++++++++++++++++------- > --- > lib/librte_hash/rte_cuckoo_hash.h | 6 +- > lib/librte_hash/rte_hash.h | 63 ++++- > lib/librte_hash/rte_hash_version.map | 7 + > 4 files changed, 393 insertions(+), 128 deletions(-) > > -- > 2.7.4
Hi Honnappa, Reply inlined: >-----Original Message----- > > Currently, reader-writer concurrency problems in rte_hash are > addressed using reader-writer locks. Use of reader-writer locks > results in following issues: > > 1) In many of the use cases for the hash table, writer threads > are running on control plane. If the writer is preempted while > holding the lock, it will block the readers for an extended period > resulting in packet drops. This problem seems to apply for platforms > with transactional memory support as well because of the algorithm > used for rte_rwlock_write_lock_tm: > > static inline void > rte_rwlock_write_lock_tm(rte_rwlock_t *rwl) > { > if (likely(rte_try_tm(&rwl->cnt))) > return; > rte_rwlock_write_lock(rwl); > } > > i.e. there is a posibility of using rte_rwlock_write_lock in > failure cases. [Wang, Yipeng] In our test, TSX failure happens very rarely on a TSX platform. But we agree that without TSX, the current rte_rwlock implementation may make the writer to hold a lock for a period of time. > 2) Reader-writer lock based solution does not address the following > issue. > rte_hash_lookup_xxx APIs return the index of the element in > the key store. Application(reader) can use that index to reference > other data structures in its scope. Because of this, the > index should not be freed till the application completes > using the index. [Wang, Yipeng] I agree on this use case. But I think we should provide new API functions for deletion to users who want this behavior, without changing the meaning of current API if that is possible. > Current code: > Cores Lookup Lookup > with add > 2 474 246 > 4 935 579 > 6 1387 1048 > 8 1766 1480 > 10 2119 1951 > 12 2546 2441 > > With this patch: > Cores Lookup Lookup > with add > 2 291 211 > 4 297 196 > 6 304 198 > 8 309 202 > 10 315 205 > 12 319 209 > [Wang, Yipeng] It would be good if you could provide the platform information on these results. Another comment is as you know we also have a new extension to rte_hash to enable extendable buckets and partial-key hashing. Thanks for the comments btw. Could you check if your lockless scheme also applies to those extensions?
> > Hi Honnappa, > > Reply inlined: Hi Yipeng, Thank you so much for reviewing. > > >-----Original Message----- > > > > Currently, reader-writer concurrency problems in rte_hash are > > addressed using reader-writer locks. Use of reader-writer locks > > results in following issues: > > > > 1) In many of the use cases for the hash table, writer threads > > are running on control plane. If the writer is preempted while > > holding the lock, it will block the readers for an extended period > > resulting in packet drops. This problem seems to apply for platforms > > with transactional memory support as well because of the algorithm > > used for rte_rwlock_write_lock_tm: > > > > static inline void > > rte_rwlock_write_lock_tm(rte_rwlock_t *rwl) > > { > > if (likely(rte_try_tm(&rwl->cnt))) > > return; > > rte_rwlock_write_lock(rwl); > > } > > > > i.e. there is a posibility of using rte_rwlock_write_lock in > > failure cases. > [Wang, Yipeng] In our test, TSX failure happens very rarely on a TSX > platform. But we agree that without TSX, the current rte_rwlock > implementation may make the writer to hold a lock for a period of time. > > > 2) Reader-writer lock based solution does not address the following > > issue. > > rte_hash_lookup_xxx APIs return the index of the element in > > the key store. Application(reader) can use that index to reference > > other data structures in its scope. Because of this, the > > index should not be freed till the application completes > > using the index. > [Wang, Yipeng] I agree on this use case. But I think we should provide new > API functions for deletion to users who want this behavior, without > changing the meaning of current API if that is possible. In the lock-free algorithm, the rte_hash_delete API will not free the index. The new API rte_hash_free will free the index. The solution for the algorithm with rw locks needs to be thought about. > > > Current code: > > Cores Lookup Lookup > > with add > > 2 474 246 > > 4 935 579 > > 6 1387 1048 > > 8 1766 1480 > > 10 2119 1951 > > 12 2546 2441 > > > > With this patch: > > Cores Lookup Lookup > > with add > > 2 291 211 > > 4 297 196 > > 6 304 198 > > 8 309 202 > > 10 315 205 > > 12 319 209 > > > [Wang, Yipeng] It would be good if you could provide the platform > information on these results. Apologies, I should have done that. The machine I am using is: Intel(R) Xeon(R) CPU E5-2660 v4 @ 2.00GHz, 64G memory. This is a hacked test case which is not upstreamed. In the case of 'Lookup with add' - I had equal number of threads calling 'rte_hash_add' and 'rte_hash_lookup'. In the case of 'Lookup' - a set of entries were added and all the threads called 'rte_hash_lookup'. Note that these are the numbers without htm. We have created another test case which I will upstream as next version of this patch. I will publish the numbers with that test case. So, you should be able to reproduce the numbers with that test case. > > Another comment is as you know we also have a new extension to rte_hash > to enable extendable buckets and partial-key hashing. Thanks for the > comments btw. Could you check if your lockless scheme also applies to > those extensions? Thank you for reminding me on this. I thought I had covered everything. On a relook, I have missed few key issues. I will reply on the other email thread.
>> >> Another comment is as you know we also have a new extension to rte_hash >> to enable extendable buckets and partial-key hashing. Thanks for the >> comments btw. Could you check if your lockless scheme also applies to >> those extensions? >Thank you for reminding me on this. I thought I had covered everything. On a relook, I have missed few key issues. I will reply on the >other email thread. [Wang, Yipeng] Hi, Honnappa, would like to rebase your V2 on the V4 patch I submitted for the ext table and partial-key hashing (http://patchwork.dpdk.org/cover/45620/). I will have V5 sent soon, so if it is in time, please rebase on V5. If you have any difficulty to do so we can work together on resolving the conflicts. Let's resolve the conflicts so that the final merge would be easier.