feat: memory consumption during keygen

Cargo.toml

@@ -116,6 +116,7 @@ crypto-bigint = { version = "=0.6.1", features = ["serde", "rand_core", "extra-s
 ctor = "=0.4.2"  # Constructor functions - HIGH RISK: Individual maintainer (mmastrac), test-only dependency
 dashmap = "=6.1.0"  # Concurrent hashmap - HIGH RISK: Individual maintainer (xacrimon), despite 156M+ downloads
 derive_more = { version = "=2.0.1", features = ["display"] }  # Derive macros for common traits - HIGH RISK: Individual maintainer (JelteF), despite 180M+ downloads
+dhat = "=0.3.3"  # Heap profiling - MEDIUM RISK: David Tolnay adjacent (rustacean), useful for memory debugging, dev-only
 enum_dispatch = "=0.3.13"  # Enum dispatch optimization - HIGH RISK: Individual maintainer (Anton Lazarev), despite 29M+ downloads
 futures = "=0.3.31"  # Async futures - LOW RISK: rust-lang team
 futures-util = "=0.3.31"  # Futures utilities - LOW RISK: rust-lang team

core/service/Cargo.toml

@@ -65,6 +65,7 @@ cfg-if.workspace = true
 ciborium.workspace = true
 clap = { workspace = true, features = ["derive"] }
 console_error_panic_hook.workspace = true
+dhat = { workspace = true, optional = true }
 enum_dispatch.workspace = true
 futures-util.workspace = true
 hex.workspace = true
@@ -183,3 +184,7 @@ insecure = [
     "threshold-fhe/testing",
     "dep:nsm-nitro-enclave-utils"
 ]
+# Memory profiling feature for debugging memory consumption issues
+# Enable with: cargo build --features memory-profiling
+# Produces dhat-heap.json file that can be viewed with https://nnethercote.github.io/dh_view/dh_view.html
+memory-profiling = ["dep:dhat"]

core/service/src/bin/kms-server.rs

@@ -331,7 +331,18 @@ async fn build_tls_config(
     Ok((server_config, client_config, verifier))
 }
 
+// Memory profiling with dhat (enable with --features memory-profiling)
+// View results at: https://nnethercote.github.io/dh_view/dh_view.html
+#[cfg(feature = "memory-profiling")]
+#[global_allocator]
+static ALLOC: dhat::Alloc = dhat::Alloc;
+
 fn main() -> anyhow::Result<()> {
+    // Initialize dhat profiler when memory-profiling feature is enabled
+    // This will produce a dhat-heap.json file on program exit
+    #[cfg(feature = "memory-profiling")]
+    let _profiler = dhat::Profiler::new_heap();
+
     let args = KmsArgs::parse();
     // NOTE: this config is only needed to set up the tokio runtime
     // we read it again in [main_exec] to set up the rest of the server

core/service/src/engine/threshold/service/key_generator.rs

@@ -53,10 +53,7 @@ use crate::{
             BaseKmsStruct, KeyGenMetadata, DSEP_PUBDATA_KEY,
         },
         keyset_configuration::InternalKeySetConfig,
-        threshold::{
-            service::{session::ImmutableSessionMaker, ThresholdFheKeys},
-            traits::KeyGenerator,
-        },
+        threshold::{service::session::ImmutableSessionMaker, traits::KeyGenerator},
         utils::MetricedError,
         validation::{
             parse_optional_proto_request_id, parse_proto_request_id, RequestIdParsingErr,
@@ -1092,39 +1089,14 @@ impl<
             }
         };
 
-        let (integer_server_key, decompression_key, sns_key) = {
-            let (
-                raw_server_key,
-                _raw_ksk_material,
-                _raw_compression_key,
-                raw_decompression_key,
-                raw_noise_squashing_key,
-                _raw_noise_squashing_compression_key,
-                _raw_rerandomization_key,
-                _raw_tag,
-            ) = pub_key_set.server_key.clone().into_raw_parts();
-            (
-                raw_server_key,
-                raw_decompression_key,
-                raw_noise_squashing_key,
-            )
-        };
-
-        let threshold_fhe_keys = ThresholdFheKeys {
-            private_keys: Arc::new(private_keys),
-            integer_server_key: Arc::new(integer_server_key),
-            sns_key: sns_key.map(Arc::new),
-            decompression_key: decompression_key.map(Arc::new),
-            meta_data: info.clone(),
-        };
-
-        //Note: We can't easily check here whether we succeeded writing to the meta store
-        //thus we can't increment the error counter if it fails
+        // Memory optimization: Storage function now serializes server_key first,
+        // then consumes it to extract components. This eliminates the need to clone the large
+        // server_key structure (~tens of GiB with production parameters).
         crypto_storage
             .write_threshold_keys_with_dkg_meta_store(
                 req_id,
                 epoch_id,
-                threshold_fhe_keys,
+                private_keys,
                 pub_key_set,
                 info,
                 meta_store,

core/service/src/engine/threshold/service/kms_impl.rs

@@ -770,5 +770,29 @@ mod tests {
 
             (priv_key_set, pub_key_set)
         }
+
+        /// Initializes dummy keys for storage tests.
+        /// Returns (PrivateKeySet, FhePubKeySet) for the new memory-optimized storage API.
+        pub fn init_dummy_for_storage<R: rand::Rng + rand::CryptoRng>(
+            param: threshold_fhe::execution::tfhe_internals::parameters::DKGParams,
+            tag: tfhe::Tag,
+            rng: &mut R,
+        ) -> (
+            PrivateKeySet<{ ResiduePolyF4Z128::EXTENSION_DEGREE }>,
+            FhePubKeySet,
+        ) {
+            let keyset = threshold_fhe::execution::tfhe_internals::test_feature::gen_key_set(
+                param, tag, rng,
+            );
+
+            let pub_key_set = FhePubKeySet {
+                public_key: keyset.public_keys.public_key,
+                server_key: keyset.public_keys.server_key,
+            };
+
+            let priv_key_set = PrivateKeySet::init_dummy(param);
+
+            (priv_key_set, pub_key_set)
+        }
     }
 }

core/service/src/engine/threshold/service/public_decryptor.rs

@@ -871,9 +871,9 @@ mod tests {
 
         let key_id = RequestId::new_random(rng);
 
-        // make a dummy private keyset
-        let (threshold_fhe_keys, fhe_key_set) =
-            ThresholdFheKeys::init_dummy(param, key_id.into(), rng);
+        // make a dummy private keyset for storage
+        let (private_keys, fhe_key_set) =
+            ThresholdFheKeys::init_dummy_for_storage(param, key_id.into(), rng);
 
         // Not a huge deal if we clone this server key since we only use small/test parameters
         tfhe::set_server_key(fhe_key_set.server_key.clone());
@@ -912,7 +912,7 @@ mod tests {
             .write_threshold_keys_with_dkg_meta_store(
                 &key_id,
                 &epoch_id,
-                threshold_fhe_keys,
+                private_keys,
                 fhe_key_set,
                 info,
                 Arc::clone(&key_meta_store),

core/service/src/engine/threshold/service/resharer.rs

@@ -5,10 +5,7 @@ use crate::{
             compute_info_standard_keygen, retrieve_parameters, BaseKmsStruct, KeyGenMetadata,
             DSEP_PUBDATA_KEY,
         },
-        threshold::{
-            service::{session::ImmutableSessionMaker, ThresholdFheKeys},
-            traits::Resharer,
-        },
+        threshold::{service::session::ImmutableSessionMaker, traits::Resharer},
         utils::MetricedError,
         validation::{
             parse_optional_proto_request_id, parse_proto_request_id, RequestIdParsingErr,
@@ -570,9 +567,6 @@ impl<PubS: Storage + Send + Sync + 'static, PrivS: StorageExt + Send + Sync + 's
             )
             .await?;
 
-            let (integer_server_key, _, _, decompression_key, sns_key, _, _, _) =
-                fhe_pubkeys.server_key.clone().into_raw_parts();
-
             // Compute all the info required for storing
             // using the same IDs and domain as we should've had the
             // DKG went through successfully
@@ -594,14 +588,6 @@ impl<PubS: Storage + Send + Sync + 'static, PrivS: StorageExt + Send + Sync + 's
                 }
             };
 
-            let threshold_fhe_keys = ThresholdFheKeys {
-                private_keys: Arc::new(new_private_key_set),
-                integer_server_key: Arc::new(integer_server_key),
-                sns_key: sns_key.map(Arc::new),
-                decompression_key: decompression_key.map(Arc::new),
-                meta_data: info.clone(),
-            };
-
             // Purge before we can overwrite, use a dummy_meta_store
             // as this was meant to update the meta store of DKG upon failing
             let dummy_meta_store = RwLock::new(MetaStore::<KeyGenMetadata>::new(1, 1));
@@ -618,12 +604,14 @@ impl<PubS: Storage + Send + Sync + 'static, PrivS: StorageExt + Send + Sync + 's
             // HOTFIX(keygen-recovery): Note that this overwrites the private storage
             // at the given key ID. It's needed as long as reshare shortcuts the
             // GW, but should be fixed long term.
+            // Memory optimization: Storage function now handles server_key
+            // component extraction internally, avoiding the need to clone server_key here.
             crypto_storage
                 .write_threshold_keys_with_reshare_meta_store(
                     &request_id,
                     &key_id_to_reshare,
                     &old_epoch_id,
-                    threshold_fhe_keys,
+                    new_private_key_set,
                     fhe_pubkeys,
                     info.clone(),
                     Arc::clone(&meta_store),

core/service/src/engine/threshold/service/user_decryptor.rs

@@ -763,9 +763,9 @@ mod tests {
             RealUserDecryptor::init_test_dummy_decryptor(base_kms, session_maker.make_immutable())
                 .await;
 
-        // make a dummy private keyset
-        let (threshold_fhe_keys, fhe_key_set) =
-            ThresholdFheKeys::init_dummy(param, key_id.into(), rng);
+        // make a dummy private keyset for storage
+        let (private_keys, fhe_key_set) =
+            ThresholdFheKeys::init_dummy_for_storage(param, key_id.into(), rng);
 
         // Not a huge deal if we clone this server key since we only use small/test parameters
         tfhe::set_server_key(fhe_key_set.server_key.clone());
@@ -794,7 +794,7 @@ mod tests {
             .write_threshold_keys_with_dkg_meta_store(
                 &key_id,
                 &epoch_id,
-                threshold_fhe_keys,
+                private_keys,
                 fhe_key_set,
                 info,
                 Arc::clone(&key_meta_store),