diff --git a/src/bench/CMakeLists.txt b/src/bench/CMakeLists.txt
index 2137beccb5a..9d40fc6fd4f 100644
--- a/src/bench/CMakeLists.txt
+++ b/src/bench/CMakeLists.txt
@@ -49,6 +49,7 @@ add_executable(bench_bitcoin
   streams_findbyte.cpp
   strencodings.cpp
   txgraph.cpp
+  txorphanage.cpp
   util_time.cpp
   verify_script.cpp
   xor.cpp
diff --git a/src/bench/txorphanage.cpp b/src/bench/txorphanage.cpp
new file mode 100644
index 00000000000..b9bb7b41c22
--- /dev/null
+++ b/src/bench/txorphanage.cpp
@@ -0,0 +1,270 @@
+// Copyright (c) The Bitcoin Core developers
+// Distributed under the MIT software license, see the accompanying
+// file COPYING or http://www.opensource.org/licenses/mit-license.php.
+
+#include <bench/bench.h>
+#include <consensus/amount.h>
+#include <net.h>
+#include <policy/policy.h>
+#include <primitives/transaction.h>
+#include <pubkey.h>
+#include <script/sign.h>
+#include <test/util/setup_common.h>
+#include <node/txorphanage.h>
+#include <util/check.h>
+#include <test/util/transaction_utils.h>
+
+#include <cstdint>
+#include <memory>
+
+static constexpr node::TxOrphanage::Usage TINY_TX_WEIGHT{240};
+static constexpr int64_t APPROX_WEIGHT_PER_INPUT{200};
+
+// Creates a transaction with num_inputs inputs and 1 output, padded to target_weight. Use this function to maximize m_outpoint_to_orphan_it operations.
+// If num_inputs is 0, we maximize the number of inputs.
+static CTransactionRef MakeTransactionBulkedTo(unsigned int num_inputs, int64_t target_weight, FastRandomContext& det_rand)
+{
+    CMutableTransaction tx;
+    assert(target_weight >= 40 + APPROX_WEIGHT_PER_INPUT);
+    if (!num_inputs) num_inputs = (target_weight - 40) / APPROX_WEIGHT_PER_INPUT;
+    for (unsigned int i = 0; i < num_inputs; ++i) {
+        tx.vin.emplace_back(Txid::FromUint256(det_rand.rand256()), 0);
+    }
+    assert(GetTransactionWeight(*MakeTransactionRef(tx)) <= target_weight);
+
+    tx.vout.resize(1);
+
+    // If necessary, pad the transaction to the target weight.
+    if (GetTransactionWeight(*MakeTransactionRef(tx)) < target_weight - 4) {
+        BulkTransaction(tx, target_weight);
+    }
+    return MakeTransactionRef(tx);
+}
+
+// Constructs a transaction using a subset of inputs[start_input : start_input + num_inputs] up to the weight_limit.
+static CTransactionRef MakeTransactionSpendingUpTo(const std::vector<CTxIn>& inputs, unsigned int start_input, unsigned int num_inputs, int64_t weight_limit)
+{
+    CMutableTransaction tx;
+    for (unsigned int i{start_input}; i < start_input + num_inputs; ++i) {
+        if (GetTransactionWeight(*MakeTransactionRef(tx)) + APPROX_WEIGHT_PER_INPUT >= weight_limit) break;
+        tx.vin.emplace_back(inputs.at(i % inputs.size()));
+    }
+    assert(tx.vin.size() > 0);
+    return MakeTransactionRef(tx);
+}
+static void OrphanageSinglePeerEviction(benchmark::Bench& bench)
+{
+    FastRandomContext det_rand{true};
+
+    // Fill up announcements slots with tiny txns, followed by a single large one
+    unsigned int NUM_TINY_TRANSACTIONS((node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE));
+
+    // Construct transactions to submit to orphanage: 1-in-1-out tiny transactions
+    std::vector<CTransactionRef> tiny_txs;
+    tiny_txs.reserve(NUM_TINY_TRANSACTIONS);
+    for (unsigned int i{0}; i < NUM_TINY_TRANSACTIONS; ++i) {
+        tiny_txs.emplace_back(MakeTransactionBulkedTo(1, TINY_TX_WEIGHT, det_rand));
+    }
+    auto large_tx = MakeTransactionBulkedTo(1, MAX_STANDARD_TX_WEIGHT, det_rand);
+    assert(GetTransactionWeight(*large_tx) <= MAX_STANDARD_TX_WEIGHT);
+
+    const auto orphanage{node::MakeTxOrphanage(/*max_global_ann=*/node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE, /*reserved_peer_usage=*/node::DEFAULT_RESERVED_ORPHAN_WEIGHT_PER_PEER)};
+
+    // Populate the orphanage. To maximize the number of evictions, first fill up with tiny transactions, then add a huge one.
+    NodeId peer{0};
+    // Add tiny transactions until we are just about to hit the memory limit, up to the max number of announcements.
+    // We use the same tiny transactions for all peers to minimize their contribution to the usage limit.
+    int64_t total_weight_to_add{0};
+    for (unsigned int txindex{0}; txindex < NUM_TINY_TRANSACTIONS; ++txindex) {
+        const auto& tx{tiny_txs.at(txindex)};
+
+        total_weight_to_add += GetTransactionWeight(*tx);
+        if (total_weight_to_add > orphanage->MaxGlobalUsage()) break;
+
+        assert(orphanage->AddTx(tx, peer));
+
+        // Sanity check: we should always be exiting at the point of hitting the weight limit.
+        assert(txindex < NUM_TINY_TRANSACTIONS - 1);
+    }
+
+    // In the real world, we always trim after each new tx.
+    // If we need to trim already, that means the benchmark is not representative of what LimitOrphans may do in a single call.
+    assert(orphanage->TotalOrphanUsage() <= orphanage->MaxGlobalUsage());
+    assert(orphanage->TotalLatencyScore() <= orphanage->MaxGlobalLatencyScore());
+    assert(orphanage->TotalOrphanUsage() + TINY_TX_WEIGHT > orphanage->MaxGlobalUsage());
+
+    bench.epochs(1).epochIterations(1).run([&]() NO_THREAD_SAFETY_ANALYSIS {
+        // Lastly, add the large transaction.
+        const auto num_announcements_before_trim{orphanage->CountAnnouncements()};
+        assert(orphanage->AddTx(large_tx, peer));
+        orphanage->LimitOrphans();
+
+        // If there are multiple peers, note that they all have the same DoS score. We will evict only 1 item at a time for each new DoSiest peer.
+        const auto num_announcements_after_trim{orphanage->CountAnnouncements()};
+        const auto num_evicted{num_announcements_before_trim - num_announcements_after_trim};
+
+        // The number of evictions is the same regardless of the number of peers. In both cases, we can exceed the
+        // usage limit using 1 maximally-sized transaction.
+        assert(num_evicted == MAX_STANDARD_TX_WEIGHT / TINY_TX_WEIGHT);
+    });
+}
+static void OrphanageMultiPeerEviction(benchmark::Bench& bench)
+{
+    // Best number is just below sqrt(DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE)
+    static constexpr unsigned int NUM_PEERS{39};
+    // All peers will have the same transactions. We want to be just under the weight limit, so divide the max usage limit by the number of unique transactions.
+    static constexpr node::TxOrphanage::Count NUM_UNIQUE_TXNS{node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE / NUM_PEERS};
+    static constexpr node::TxOrphanage::Usage TOTAL_USAGE_LIMIT{node::DEFAULT_RESERVED_ORPHAN_WEIGHT_PER_PEER * NUM_PEERS};
+    // Subtract 4 because BulkTransaction rounds up and we must avoid going over the weight limit early.
+    static constexpr node::TxOrphanage::Usage LARGE_TX_WEIGHT{TOTAL_USAGE_LIMIT / NUM_UNIQUE_TXNS - 4};
+    static_assert(LARGE_TX_WEIGHT >= TINY_TX_WEIGHT * 2, "Tx is too small, increase NUM_PEERS");
+    // The orphanage does not permit any transactions larger than 400'000, so this test will not work if the large tx is much larger.
+    static_assert(LARGE_TX_WEIGHT <= MAX_STANDARD_TX_WEIGHT, "Tx is too large, decrease NUM_PEERS");
+
+    FastRandomContext det_rand{true};
+    // Construct large transactions
+    std::vector<CTransactionRef> shared_txs;
+    shared_txs.reserve(NUM_UNIQUE_TXNS);
+    for (unsigned int i{0}; i < NUM_UNIQUE_TXNS; ++i) {
+        shared_txs.emplace_back(MakeTransactionBulkedTo(9, LARGE_TX_WEIGHT, det_rand));
+    }
+    std::vector<size_t> indexes;
+    indexes.resize(NUM_UNIQUE_TXNS);
+    std::iota(indexes.begin(), indexes.end(), 0);
+
+    const auto orphanage{node::MakeTxOrphanage(/*max_global_ann=*/node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE, /*reserved_peer_usage=*/node::DEFAULT_RESERVED_ORPHAN_WEIGHT_PER_PEER)};
+    // Every peer sends the same transactions, all from shared_txs.
+    // Each peer has 1 or 2 assigned transactions, which they must place as the last and second-to-last positions.
+    // The assignments ensure that every transaction is in some peer's last 2 transactions, and is thus remains in the orphanage until the end of LimitOrphans.
+    static_assert(NUM_UNIQUE_TXNS <= NUM_PEERS * 2);
+
+    // We need each peer to send some transactions so that the global limit (which is a function of the number of peers providing at least 1 announcement) rises.
+    for (unsigned int i{0}; i < NUM_UNIQUE_TXNS; ++i) {
+        for (NodeId peer{0}; peer < NUM_PEERS; ++peer) {
+            const CTransactionRef& reserved_last_tx{shared_txs.at(peer)};
+            CTransactionRef reserved_second_to_last_tx{peer < NUM_UNIQUE_TXNS - NUM_PEERS ? shared_txs.at(peer + NUM_PEERS) : nullptr};
+
+            const auto& tx{shared_txs.at(indexes.at(i))};
+            if (tx == reserved_last_tx) {
+                // Skip
+            } else if (reserved_second_to_last_tx && tx == reserved_second_to_last_tx) {
+                // Skip
+            } else {
+                orphanage->AddTx(tx, peer);
+            }
+        }
+    }
+
+    // Now add the final reserved transactions.
+    for (NodeId peer{0}; peer < NUM_PEERS; ++peer) {
+        const CTransactionRef& reserved_last_tx{shared_txs.at(peer)};
+        CTransactionRef reserved_second_to_last_tx{peer < NUM_UNIQUE_TXNS - NUM_PEERS ? shared_txs.at(peer + NUM_PEERS) : nullptr};
+        // Add the final reserved transactions.
+        if (reserved_second_to_last_tx) {
+            orphanage->AddTx(reserved_second_to_last_tx, peer);
+        }
+        orphanage->AddTx(reserved_last_tx, peer);
+    }
+
+    assert(orphanage->CountAnnouncements() == NUM_PEERS * NUM_UNIQUE_TXNS);
+    const auto total_usage{orphanage->TotalOrphanUsage()};
+    const auto max_usage{orphanage->MaxGlobalUsage()};
+    assert(max_usage - total_usage <= LARGE_TX_WEIGHT);
+    assert(orphanage->TotalLatencyScore() <= orphanage->MaxGlobalLatencyScore());
+
+    auto last_tx = MakeTransactionBulkedTo(0, max_usage - total_usage + 1, det_rand);
+
+    bench.epochs(1).epochIterations(1).run([&]() NO_THREAD_SAFETY_ANALYSIS {
+        const auto num_announcements_before_trim{orphanage->CountAnnouncements()};
+        // There is a small gap between the total usage and the max usage. Add a transaction to fill it.
+        assert(orphanage->AddTx(last_tx, 0));
+        orphanage->LimitOrphans();
+
+        // If there are multiple peers, note that they all have the same DoS score. We will evict only 1 item at a time for each new DoSiest peer.
+        const auto num_evicted{num_announcements_before_trim - orphanage->CountAnnouncements() + 1};
+        // The trimming happens as a round robin. In the first NUM_UNIQUE_TXNS - 2 rounds for each peer, only duplicates are evicted.
+        // Once each peer has 2 transactions left, it's possible to select a peer whose oldest transaction is unique.
+        assert(num_evicted >= (NUM_UNIQUE_TXNS - 2) * NUM_PEERS);
+    });
+}
+
+static void OrphanageEraseAll(benchmark::Bench& bench, bool block_or_disconnect)
+{
+    FastRandomContext det_rand{true};
+    const auto orphanage{node::MakeTxOrphanage(/*max_global_ann=*/node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE, /*reserved_peer_usage=*/node::DEFAULT_RESERVED_ORPHAN_WEIGHT_PER_PEER)};
+    // This is an unrealistically large number of inputs for a block, as there is almost no room given to witness data,
+    // outputs, and overhead for individual transactions. The entire block is 1 transaction with 20,000 inputs.
+    constexpr unsigned int NUM_BLOCK_INPUTS{MAX_BLOCK_WEIGHT / APPROX_WEIGHT_PER_INPUT};
+    const auto block_tx{MakeTransactionBulkedTo(NUM_BLOCK_INPUTS, MAX_BLOCK_WEIGHT - 4000, det_rand)};
+    CBlock block;
+    block.vtx.push_back(block_tx);
+
+    // Transactions with 9 inputs maximize the computation / LatencyScore ratio.
+    constexpr unsigned int INPUTS_PER_TX{9};
+    constexpr unsigned int NUM_PEERS{125};
+    constexpr unsigned int NUM_TXNS_PER_PEER = node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE / NUM_PEERS;
+
+    // Divide the block's inputs evenly among the peers.
+    constexpr unsigned int INPUTS_PER_PEER = NUM_BLOCK_INPUTS / NUM_PEERS;
+    static_assert(INPUTS_PER_PEER > 0);
+    // All the block inputs are spent by the orphanage transactions. Each peer is assigned 76 of them.
+    // Each peer has 24 transactions spending 9 inputs each, so jumping by 3 ensures we cover all of the inputs.
+    static_assert(7 * NUM_TXNS_PER_PEER + INPUTS_PER_TX - 1 >= INPUTS_PER_PEER);
+
+    for (NodeId peer{0}; peer < NUM_PEERS; ++peer) {
+        int64_t weight_left_for_peer{node::DEFAULT_RESERVED_ORPHAN_WEIGHT_PER_PEER};
+        for (unsigned int txnum{0}; txnum < node::DEFAULT_MAX_ORPHANAGE_LATENCY_SCORE / NUM_PEERS; ++txnum) {
+            // Transactions must be unique since they use different (though overlapping) inputs.
+            const unsigned int start_input = peer * INPUTS_PER_PEER + txnum * 7;
+
+            // Note that we shouldn't be able to hit the weight limit with these small transactions.
+            const int64_t weight_limit{std::min<int64_t>(weight_left_for_peer, MAX_STANDARD_TX_WEIGHT)};
+            auto ptx = MakeTransactionSpendingUpTo(block_tx->vin, /*start_input=*/start_input, /*num_inputs=*/INPUTS_PER_TX, /*weight_limit=*/weight_limit);
+
+            assert(GetTransactionWeight(*ptx) <= MAX_STANDARD_TX_WEIGHT);
+            assert(!orphanage->HaveTx(ptx->GetWitnessHash()));
+            assert(orphanage->AddTx(ptx, peer));
+
+            weight_left_for_peer -= GetTransactionWeight(*ptx);
+            if (weight_left_for_peer < TINY_TX_WEIGHT * 2) break;
+        }
+    }
+
+    // If these fail, it means this benchmark is not realistic because the orphanage would have been trimmed already.
+    assert(orphanage->TotalLatencyScore() <= orphanage->MaxGlobalLatencyScore());
+    assert(orphanage->TotalOrphanUsage() <= orphanage->MaxGlobalUsage());
+
+    // 3000 announcements (and unique transactions) in the orphanage.
+    // They spend a total of 27,000 inputs (20,000 unique ones)
+    assert(orphanage->CountAnnouncements() == NUM_PEERS * NUM_TXNS_PER_PEER);
+    assert(orphanage->TotalLatencyScore() == orphanage->CountAnnouncements());
+
+    bench.epochs(1).epochIterations(1).run([&]() NO_THREAD_SAFETY_ANALYSIS {
+        if (block_or_disconnect) {
+            // Erase everything through EraseForBlock.
+            // Every tx conflicts with this block.
+            orphanage->EraseForBlock(block);
+            assert(orphanage->CountAnnouncements() == 0);
+        } else {
+            // Erase everything through EraseForPeer.
+            for (NodeId peer{0}; peer < NUM_PEERS; ++peer) {
+                orphanage->EraseForPeer(peer);
+            }
+            assert(orphanage->CountAnnouncements() == 0);
+        }
+    });
+}
+
+static void OrphanageEraseForBlock(benchmark::Bench& bench)
+{
+    OrphanageEraseAll(bench, /*block_or_disconnect=*/true);
+}
+static void OrphanageEraseForPeer(benchmark::Bench& bench)
+{
+    OrphanageEraseAll(bench, /*block_or_disconnect=*/false);
+}
+
+BENCHMARK(OrphanageSinglePeerEviction, benchmark::PriorityLevel::LOW);
+BENCHMARK(OrphanageMultiPeerEviction, benchmark::PriorityLevel::LOW);
+BENCHMARK(OrphanageEraseForBlock, benchmark::PriorityLevel::LOW);
+BENCHMARK(OrphanageEraseForPeer, benchmark::PriorityLevel::LOW);