#include <bits/stdc++.h>
using namespace std;
using lint = long long;
using pint = pair<int, int>;
using plint = pair<lint, lint>;
struct fast_ios {
    fast_ios() { cin.tie(nullptr), ios::sync_with_stdio(false), cout << fixed << setprecision(20); };
} fast_ios_;
#define FOR(i, begin, end) for (int i = (begin), i##_end_ = (end); i < i##_end_; i++)
#define IFOR(i, begin, end) for (int i = (end)-1, i##_begin_ = (begin); i >= i##_begin_; i--)
#define REP(i, n) FOR(i, 0, n)
#define IREP(i, n) IFOR(i, 0, n)
#define ALL(x) (x).begin(), (x).end()
//
template <typename T, typename V>
void ndarray(vector<T>& vec, const V& val, int len) { vec.assign(len, val); }
template <typename T, typename V, typename... Args>
void ndarray(vector<T>& vec, const V& val, int len, Args... args)
{
    vec.resize(len), for_each(begin(vec), end(vec), [&](T& v) { ndarray(v, val, args...); });
}
template <typename T>
bool chmax(T& m, const T q) { return m < q ? (m = q, true) : false; }
template <typename T>
bool chmin(T& m, const T q) { return m > q ? (m = q, true) : false; }
template <typename T1, typename T2>
pair<T1, T2> operator+(const pair<T1, T2>& l, const pair<T1, T2>& r) { return make_pair(l.first + r.first, l.second + r.second); }
template <typename T1, typename T2>
pair<T1, T2> operator-(const pair<T1, T2>& l, const pair<T1, T2>& r) { return make_pair(l.first - r.first, l.second - r.second); }
template <typename T>
vector<T> srtunq(vector<T> vec) { return sort(vec.begin(), vec.end()), vec.erase(unique(vec.begin(), vec.end()), vec.end()), vec; }
template <typename T>
istream& operator>>(istream& is, vector<T>& vec)
{
    return for_each(begin(vec), end(vec), [&](T& v) { is >> v; }), is;
}

// output
template <typename T, typename V>
ostream& dmpseq(ostream&, const T&, const string&, const string&, const string&);
#if __cplusplus >= 201703L
template <typename... T>
ostream& operator<<(ostream& os, const tuple<T...>& tpl)
{
    return apply([&os](auto&&... args) { ((os << args << ','), ...); }, tpl), os;
}
#endif
//
template <typename T1, typename T2>
ostream& operator<<(ostream& os, const pair<T1, T2>& p) { return os << '(' << p.first << ',' << p.second << ')'; }
template <typename T>
ostream& operator<<(ostream& os, const vector<T>& x) { return dmpseq<vector<T>, T>(os, x, "[", ",", "]"); }
template <typename T>
ostream& operator<<(ostream& os, const deque<T>& x) { return dmpseq<deque<T>, T>(os, x, "deq[", ",", "]"); }
template <typename T>
ostream& operator<<(ostream& os, const set<T>& x) { return dmpseq<set<T>, T>(os, x, "{", ",", "}"); }
template <typename T, typename TH>
ostream& operator<<(ostream& os, const unordered_set<T, TH>& x) { return dmpseq<unordered_set<T, TH>, T>(os, x, "{", ",", "}"); }
template <typename T>
ostream& operator<<(ostream& os, const multiset<T>& x) { return dmpseq<multiset<T>, T>(os, x, "{", ",", "}"); }
template <typename TK, typename T>
ostream& operator<<(ostream& os, const map<TK, T>& x) { return dmpseq<map<TK, T>, pair<TK, T>>(os, x, "{", ",", "}"); }
template <typename TK, typename T, typename TH>
ostream& operator<<(ostream& os, const unordered_map<TK, T, TH>& x) { return dmpseq<unordered_map<TK, T, TH>, pair<TK, T>>(os, x, "{", ",", "}"); }
template <typename T, typename V>
ostream& dmpseq(ostream& os, const T& seq, const string& pre, const string& sp, const string& suf)
{
    return os << pre, for_each(begin(seq), end(seq), [&](V x) { os << x << sp; }), os << suf;
}
template <typename T>
void print(const vector<T>& x) { dmpseq<vector<T>, T>(cout, x, "", " ", "\n"); }
#ifdef HITONANODE_LOCAL
#define dbg(x) cerr << #x << " = " << (x) << " (L" << __LINE__ << ") " << __FILE__ << endl
#else
#define dbg(x) {}
#endif

template <typename T>
struct BIT : std::vector<T>
{
    BIT(int len = 0) : std::vector<T>(len + 1) {}
    void reset() { fill(this->begin(), this->end(), 0); }
    void add(int pos, T v) {
        while (pos > 0 and pos < (int)this->size()) (*this)[pos] += v, pos += pos & -pos;
    }
    T sum(int pos) const { // (0, pos]
        T res = 0;
        while (pos > 0) res += (*this)[pos], pos -= pos & -pos;
        return res;
    }
    friend std::ostream &operator<<(std::ostream &os, const BIT &bit) {
        T prv = 0;
        os << '[';
        for (int i = 1; i < (int)bit.size(); i++) {
            T now = bit.sum(i);
            os << now - prv << ",";
            prv = now;
        }
        os << ']';
        return os;
    }
};

#include <ext/pb_ds/assoc_container.hpp>
#include <ext/pb_ds/tree_policy.hpp>
#include <ext/pb_ds/tag_and_trait.hpp>
using namespace __gnu_pbds; // find_by_order(), order_of_key()
template <typename TK> using pbds_set = tree<TK, null_type, less<TK>, rb_tree_tag, tree_order_statistics_node_update>;
template <typename TK, typename TV> using pbds_map = tree<TK, TV, less<TK>, rb_tree_tag, tree_order_statistics_node_update>;


int main()
{
    int N, K;
    cin >> N >> K;
    vector<lint> A(N);
    cin >> A;
    vector<plint> A2I(N);
    REP(i, N) A2I[i] = plint(A[i], i);
    sort(ALL(A2I));
    vector<int> ii(N);
    REP(i, N) ii[A2I[i].second] = i + 1;
    BIT<lint> bit(N + 10);

    lint ret = A[K - 1];

    set<pint> cur_set;
    set<pint> rightset;
    pbds_set<pint> all_set;
    FOR(i, K - 1, N) all_set.insert(pint(A[i], i)), bit.add(ii[i], A[i]);
    FOR(i, K - 1, 2 * (K - 1) + 1) if (i < N) cur_set.emplace(A[i], i);
    FOR(i, 2 * K - 1, N) rightset.emplace(A[i], i);

    for (lint m = 2; m * K <= N; m++) {
        if (cur_set.empty() or rightset.empty()) break;
        dbg(m);
        dbg(rightset);
        dbg(cur_set);
        lint cum = rightset.begin()->first + cur_set.begin()->first;
        int ir = rightset.begin()->second, ic = cur_set.begin()->second;
        all_set.erase(pint(A[ir], ir));
        all_set.erase(pint(A[ic], ic));
        bit.add(ii[ir], -A[ir]);
        bit.add(ii[ic], -A[ic]);
        dbg(cum);
        dbg(bit);
        if (int(all_set.size()) < m - 2) break;
        if (int(all_set.size()) > m - 3 and m >= 3)
        {
            auto p = *all_set.find_by_order(m - 3);
            cum += bit.sum(ii[p.second]);
        }
        chmin(ret, cum);

        all_set.insert(pint(A[ir], ir));
        all_set.insert(pint(A[ic], ic));
        bit.add(ii[ir], A[ir]);
        bit.add(ii[ic], A[ic]);

        const int l = (K - 1) * (m - 1);
        FOR(i, l, l + K - 1)
        {
            if (i >= N) break;
            bit.add(ii[i], -A[i]);
            cur_set.erase(pint(A[i], i));
            all_set.erase(pint(A[i], i));
        }
        FOR(i, l + K, l + K * 2 - 1) if (i < N)
        {
            cur_set.insert(pint(A[i], i));
        }
        FOR(i, m * K - 1, (m + 1) * K - 1)
        {
            if (i >= N) break;
            rightset.erase(pint(A[i], i));
        }
    }
    dbg(cur_set);
    dbg(rightset);
    dbg(bit);
    cout << ret << '\n';
}